| 1 | //===--- CGStmtOpenMP.cpp - Emit LLVM Code from Statements ----------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This contains code to emit OpenMP nodes as LLVM code. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CGCleanup.h" |
| 14 | #include "CGOpenMPRuntime.h" |
| 15 | #include "CodeGenFunction.h" |
| 16 | #include "CodeGenModule.h" |
| 17 | #include "TargetInfo.h" |
| 18 | #include "clang/AST/ASTContext.h" |
| 19 | #include "clang/AST/Attr.h" |
| 20 | #include "clang/AST/DeclOpenMP.h" |
| 21 | #include "clang/AST/OpenMPClause.h" |
| 22 | #include "clang/AST/Stmt.h" |
| 23 | #include "clang/AST/StmtOpenMP.h" |
| 24 | #include "clang/AST/StmtVisitor.h" |
| 25 | #include "clang/Basic/OpenMPKinds.h" |
| 26 | #include "clang/Basic/PrettyStackTrace.h" |
| 27 | #include "llvm/Frontend/OpenMP/OMPConstants.h" |
| 28 | #include "llvm/Frontend/OpenMP/OMPIRBuilder.h" |
| 29 | #include "llvm/IR/Constants.h" |
| 30 | #include "llvm/IR/Instructions.h" |
| 31 | #include "llvm/Support/AtomicOrdering.h" |
| 32 | using namespace clang; |
| 33 | using namespace CodeGen; |
| 34 | using namespace llvm::omp; |
| 35 | |
| 36 | static const VarDecl *getBaseDecl(const Expr *Ref); |
| 37 | |
| 38 | namespace { |
| 39 | /// Lexical scope for OpenMP executable constructs, that handles correct codegen |
| 40 | /// for captured expressions. |
| 41 | class OMPLexicalScope : public CodeGenFunction::LexicalScope { |
| 42 | void emitPreInitStmt(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 43 | for (const auto *C : S.clauses()) { |
| 44 | if (const auto *CPI = OMPClauseWithPreInit::get(C)) { |
| 45 | if (const auto *PreInit = |
| 46 | cast_or_null<DeclStmt>(CPI->getPreInitStmt())) { |
| 47 | for (const auto *I : PreInit->decls()) { |
| 48 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { |
| 49 | CGF.EmitVarDecl(cast<VarDecl>(*I)); |
| 50 | } else { |
| 51 | CodeGenFunction::AutoVarEmission Emission = |
| 52 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); |
| 53 | CGF.EmitAutoVarCleanups(Emission); |
| 54 | } |
| 55 | } |
| 56 | } |
| 57 | } |
| 58 | } |
| 59 | } |
| 60 | CodeGenFunction::OMPPrivateScope InlinedShareds; |
| 61 | |
| 62 | static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) { |
| 63 | return CGF.LambdaCaptureFields.lookup(VD) || |
| 64 | (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) || |
| 65 | (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl) && |
| 66 | cast<BlockDecl>(CGF.CurCodeDecl)->capturesVariable(VD)); |
| 67 | } |
| 68 | |
| 69 | public: |
| 70 | OMPLexicalScope( |
| 71 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 72 | const llvm::Optional<OpenMPDirectiveKind> CapturedRegion = llvm::None, |
| 73 | const bool EmitPreInitStmt = true) |
| 74 | : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()), |
| 75 | InlinedShareds(CGF) { |
| 76 | if (EmitPreInitStmt) |
| 77 | emitPreInitStmt(CGF, S); |
| 78 | if (!CapturedRegion.hasValue()) |
| 79 | return; |
| 80 | assert(S.hasAssociatedStmt() && |
| 81 | "Expected associated statement for inlined directive." ); |
| 82 | const CapturedStmt *CS = S.getCapturedStmt(*CapturedRegion); |
| 83 | for (const auto &C : CS->captures()) { |
| 84 | if (C.capturesVariable() || C.capturesVariableByCopy()) { |
| 85 | auto *VD = C.getCapturedVar(); |
| 86 | assert(VD == VD->getCanonicalDecl() && |
| 87 | "Canonical decl must be captured." ); |
| 88 | DeclRefExpr DRE( |
| 89 | CGF.getContext(), const_cast<VarDecl *>(VD), |
| 90 | isCapturedVar(CGF, VD) || (CGF.CapturedStmtInfo && |
| 91 | InlinedShareds.isGlobalVarCaptured(VD)), |
| 92 | VD->getType().getNonReferenceType(), VK_LValue, C.getLocation()); |
| 93 | InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address { |
| 94 | return CGF.EmitLValue(&DRE).getAddress(CGF); |
| 95 | }); |
| 96 | } |
| 97 | } |
| 98 | (void)InlinedShareds.Privatize(); |
| 99 | } |
| 100 | }; |
| 101 | |
| 102 | /// Lexical scope for OpenMP parallel construct, that handles correct codegen |
| 103 | /// for captured expressions. |
| 104 | class OMPParallelScope final : public OMPLexicalScope { |
| 105 | bool EmitPreInitStmt(const OMPExecutableDirective &S) { |
| 106 | OpenMPDirectiveKind Kind = S.getDirectiveKind(); |
| 107 | return !(isOpenMPTargetExecutionDirective(Kind) || |
| 108 | isOpenMPLoopBoundSharingDirective(Kind)) && |
| 109 | isOpenMPParallelDirective(Kind); |
| 110 | } |
| 111 | |
| 112 | public: |
| 113 | OMPParallelScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 114 | : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None, |
| 115 | EmitPreInitStmt(S)) {} |
| 116 | }; |
| 117 | |
| 118 | /// Lexical scope for OpenMP teams construct, that handles correct codegen |
| 119 | /// for captured expressions. |
| 120 | class OMPTeamsScope final : public OMPLexicalScope { |
| 121 | bool EmitPreInitStmt(const OMPExecutableDirective &S) { |
| 122 | OpenMPDirectiveKind Kind = S.getDirectiveKind(); |
| 123 | return !isOpenMPTargetExecutionDirective(Kind) && |
| 124 | isOpenMPTeamsDirective(Kind); |
| 125 | } |
| 126 | |
| 127 | public: |
| 128 | OMPTeamsScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 129 | : OMPLexicalScope(CGF, S, /*CapturedRegion=*/llvm::None, |
| 130 | EmitPreInitStmt(S)) {} |
| 131 | }; |
| 132 | |
| 133 | /// Private scope for OpenMP loop-based directives, that supports capturing |
| 134 | /// of used expression from loop statement. |
| 135 | class OMPLoopScope : public CodeGenFunction::RunCleanupsScope { |
| 136 | void emitPreInitStmt(CodeGenFunction &CGF, const OMPLoopDirective &S) { |
| 137 | CodeGenFunction::OMPMapVars PreCondVars; |
| 138 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 139 | for (const auto *E : S.counters()) { |
| 140 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 141 | EmittedAsPrivate.insert(VD->getCanonicalDecl()); |
| 142 | (void)PreCondVars.setVarAddr( |
| 143 | CGF, VD, CGF.CreateMemTemp(VD->getType().getNonReferenceType())); |
| 144 | } |
| 145 | // Mark private vars as undefs. |
| 146 | for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) { |
| 147 | for (const Expr *IRef : C->varlists()) { |
| 148 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(IRef)->getDecl()); |
| 149 | if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 150 | (void)PreCondVars.setVarAddr( |
| 151 | CGF, OrigVD, |
| 152 | Address(llvm::UndefValue::get( |
| 153 | CGF.ConvertTypeForMem(CGF.getContext().getPointerType( |
| 154 | OrigVD->getType().getNonReferenceType()))), |
| 155 | CGF.getContext().getDeclAlign(OrigVD))); |
| 156 | } |
| 157 | } |
| 158 | } |
| 159 | (void)PreCondVars.apply(CGF); |
| 160 | // Emit init, __range and __end variables for C++ range loops. |
| 161 | const Stmt *Body = |
| 162 | S.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(); |
| 163 | for (unsigned Cnt = 0; Cnt < S.getCollapsedNumber(); ++Cnt) { |
| 164 | Body = OMPLoopDirective::tryToFindNextInnerLoop( |
| 165 | Body, /*TryImperfectlyNestedLoops=*/true); |
| 166 | if (auto *For = dyn_cast<ForStmt>(Body)) { |
| 167 | Body = For->getBody(); |
| 168 | } else { |
| 169 | assert(isa<CXXForRangeStmt>(Body) && |
| 170 | "Expected canonical for loop or range-based for loop." ); |
| 171 | auto *CXXFor = cast<CXXForRangeStmt>(Body); |
| 172 | if (const Stmt *Init = CXXFor->getInit()) |
| 173 | CGF.EmitStmt(Init); |
| 174 | CGF.EmitStmt(CXXFor->getRangeStmt()); |
| 175 | CGF.EmitStmt(CXXFor->getEndStmt()); |
| 176 | Body = CXXFor->getBody(); |
| 177 | } |
| 178 | } |
| 179 | if (const auto *PreInits = cast_or_null<DeclStmt>(S.getPreInits())) { |
| 180 | for (const auto *I : PreInits->decls()) |
| 181 | CGF.EmitVarDecl(cast<VarDecl>(*I)); |
| 182 | } |
| 183 | PreCondVars.restore(CGF); |
| 184 | } |
| 185 | |
| 186 | public: |
| 187 | OMPLoopScope(CodeGenFunction &CGF, const OMPLoopDirective &S) |
| 188 | : CodeGenFunction::RunCleanupsScope(CGF) { |
| 189 | emitPreInitStmt(CGF, S); |
| 190 | } |
| 191 | }; |
| 192 | |
| 193 | class OMPSimdLexicalScope : public CodeGenFunction::LexicalScope { |
| 194 | CodeGenFunction::OMPPrivateScope InlinedShareds; |
| 195 | |
| 196 | static bool isCapturedVar(CodeGenFunction &CGF, const VarDecl *VD) { |
| 197 | return CGF.LambdaCaptureFields.lookup(VD) || |
| 198 | (CGF.CapturedStmtInfo && CGF.CapturedStmtInfo->lookup(VD)) || |
| 199 | (CGF.CurCodeDecl && isa<BlockDecl>(CGF.CurCodeDecl) && |
| 200 | cast<BlockDecl>(CGF.CurCodeDecl)->capturesVariable(VD)); |
| 201 | } |
| 202 | |
| 203 | public: |
| 204 | OMPSimdLexicalScope(CodeGenFunction &CGF, const OMPExecutableDirective &S) |
| 205 | : CodeGenFunction::LexicalScope(CGF, S.getSourceRange()), |
| 206 | InlinedShareds(CGF) { |
| 207 | for (const auto *C : S.clauses()) { |
| 208 | if (const auto *CPI = OMPClauseWithPreInit::get(C)) { |
| 209 | if (const auto *PreInit = |
| 210 | cast_or_null<DeclStmt>(CPI->getPreInitStmt())) { |
| 211 | for (const auto *I : PreInit->decls()) { |
| 212 | if (!I->hasAttr<OMPCaptureNoInitAttr>()) { |
| 213 | CGF.EmitVarDecl(cast<VarDecl>(*I)); |
| 214 | } else { |
| 215 | CodeGenFunction::AutoVarEmission Emission = |
| 216 | CGF.EmitAutoVarAlloca(cast<VarDecl>(*I)); |
| 217 | CGF.EmitAutoVarCleanups(Emission); |
| 218 | } |
| 219 | } |
| 220 | } |
| 221 | } else if (const auto *UDP = dyn_cast<OMPUseDevicePtrClause>(C)) { |
| 222 | for (const Expr *E : UDP->varlists()) { |
| 223 | const Decl *D = cast<DeclRefExpr>(E)->getDecl(); |
| 224 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(D)) |
| 225 | CGF.EmitVarDecl(*OED); |
| 226 | } |
| 227 | } else if (const auto *UDP = dyn_cast<OMPUseDeviceAddrClause>(C)) { |
| 228 | for (const Expr *E : UDP->varlists()) { |
| 229 | const Decl *D = getBaseDecl(E); |
| 230 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(D)) |
| 231 | CGF.EmitVarDecl(*OED); |
| 232 | } |
| 233 | } |
| 234 | } |
| 235 | if (!isOpenMPSimdDirective(S.getDirectiveKind())) |
| 236 | CGF.EmitOMPPrivateClause(S, InlinedShareds); |
| 237 | if (const auto *TG = dyn_cast<OMPTaskgroupDirective>(&S)) { |
| 238 | if (const Expr *E = TG->getReductionRef()) |
| 239 | CGF.EmitVarDecl(*cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl())); |
| 240 | } |
| 241 | const auto *CS = cast_or_null<CapturedStmt>(S.getAssociatedStmt()); |
| 242 | while (CS) { |
| 243 | for (auto &C : CS->captures()) { |
| 244 | if (C.capturesVariable() || C.capturesVariableByCopy()) { |
| 245 | auto *VD = C.getCapturedVar(); |
| 246 | assert(VD == VD->getCanonicalDecl() && |
| 247 | "Canonical decl must be captured." ); |
| 248 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(VD), |
| 249 | isCapturedVar(CGF, VD) || |
| 250 | (CGF.CapturedStmtInfo && |
| 251 | InlinedShareds.isGlobalVarCaptured(VD)), |
| 252 | VD->getType().getNonReferenceType(), VK_LValue, |
| 253 | C.getLocation()); |
| 254 | InlinedShareds.addPrivate(VD, [&CGF, &DRE]() -> Address { |
| 255 | return CGF.EmitLValue(&DRE).getAddress(CGF); |
| 256 | }); |
| 257 | } |
| 258 | } |
| 259 | CS = dyn_cast<CapturedStmt>(CS->getCapturedStmt()); |
| 260 | } |
| 261 | (void)InlinedShareds.Privatize(); |
| 262 | } |
| 263 | }; |
| 264 | |
| 265 | } // namespace |
| 266 | |
| 267 | static void emitCommonOMPTargetDirective(CodeGenFunction &CGF, |
| 268 | const OMPExecutableDirective &S, |
| 269 | const RegionCodeGenTy &CodeGen); |
| 270 | |
| 271 | LValue CodeGenFunction::EmitOMPSharedLValue(const Expr *E) { |
| 272 | if (const auto *OrigDRE = dyn_cast<DeclRefExpr>(E)) { |
| 273 | if (const auto *OrigVD = dyn_cast<VarDecl>(OrigDRE->getDecl())) { |
| 274 | OrigVD = OrigVD->getCanonicalDecl(); |
| 275 | bool IsCaptured = |
| 276 | LambdaCaptureFields.lookup(OrigVD) || |
| 277 | (CapturedStmtInfo && CapturedStmtInfo->lookup(OrigVD)) || |
| 278 | (CurCodeDecl && isa<BlockDecl>(CurCodeDecl)); |
| 279 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), IsCaptured, |
| 280 | OrigDRE->getType(), VK_LValue, OrigDRE->getExprLoc()); |
| 281 | return EmitLValue(&DRE); |
| 282 | } |
| 283 | } |
| 284 | return EmitLValue(E); |
| 285 | } |
| 286 | |
| 287 | llvm::Value *CodeGenFunction::getTypeSize(QualType Ty) { |
| 288 | ASTContext &C = getContext(); |
| 289 | llvm::Value *Size = nullptr; |
| 290 | auto SizeInChars = C.getTypeSizeInChars(Ty); |
| 291 | if (SizeInChars.isZero()) { |
| 292 | // getTypeSizeInChars() returns 0 for a VLA. |
| 293 | while (const VariableArrayType *VAT = C.getAsVariableArrayType(Ty)) { |
| 294 | VlaSizePair VlaSize = getVLASize(VAT); |
| 295 | Ty = VlaSize.Type; |
| 296 | Size = Size ? Builder.CreateNUWMul(Size, VlaSize.NumElts) |
| 297 | : VlaSize.NumElts; |
| 298 | } |
| 299 | SizeInChars = C.getTypeSizeInChars(Ty); |
| 300 | if (SizeInChars.isZero()) |
| 301 | return llvm::ConstantInt::get(SizeTy, /*V=*/0); |
| 302 | return Builder.CreateNUWMul(Size, CGM.getSize(SizeInChars)); |
| 303 | } |
| 304 | return CGM.getSize(SizeInChars); |
| 305 | } |
| 306 | |
| 307 | void CodeGenFunction::GenerateOpenMPCapturedVars( |
| 308 | const CapturedStmt &S, SmallVectorImpl<llvm::Value *> &CapturedVars) { |
| 309 | const RecordDecl *RD = S.getCapturedRecordDecl(); |
| 310 | auto CurField = RD->field_begin(); |
| 311 | auto CurCap = S.captures().begin(); |
| 312 | for (CapturedStmt::const_capture_init_iterator I = S.capture_init_begin(), |
| 313 | E = S.capture_init_end(); |
| 314 | I != E; ++I, ++CurField, ++CurCap) { |
| 315 | if (CurField->hasCapturedVLAType()) { |
| 316 | const VariableArrayType *VAT = CurField->getCapturedVLAType(); |
| 317 | llvm::Value *Val = VLASizeMap[VAT->getSizeExpr()]; |
| 318 | CapturedVars.push_back(Val); |
| 319 | } else if (CurCap->capturesThis()) { |
| 320 | CapturedVars.push_back(CXXThisValue); |
| 321 | } else if (CurCap->capturesVariableByCopy()) { |
| 322 | llvm::Value *CV = EmitLoadOfScalar(EmitLValue(*I), CurCap->getLocation()); |
| 323 | |
| 324 | // If the field is not a pointer, we need to save the actual value |
| 325 | // and load it as a void pointer. |
| 326 | if (!CurField->getType()->isAnyPointerType()) { |
| 327 | ASTContext &Ctx = getContext(); |
| 328 | Address DstAddr = CreateMemTemp( |
| 329 | Ctx.getUIntPtrType(), |
| 330 | Twine(CurCap->getCapturedVar()->getName(), ".casted" )); |
| 331 | LValue DstLV = MakeAddrLValue(DstAddr, Ctx.getUIntPtrType()); |
| 332 | |
| 333 | llvm::Value *SrcAddrVal = EmitScalarConversion( |
| 334 | DstAddr.getPointer(), Ctx.getPointerType(Ctx.getUIntPtrType()), |
| 335 | Ctx.getPointerType(CurField->getType()), CurCap->getLocation()); |
| 336 | LValue SrcLV = |
| 337 | MakeNaturalAlignAddrLValue(SrcAddrVal, CurField->getType()); |
| 338 | |
| 339 | // Store the value using the source type pointer. |
| 340 | EmitStoreThroughLValue(RValue::get(CV), SrcLV); |
| 341 | |
| 342 | // Load the value using the destination type pointer. |
| 343 | CV = EmitLoadOfScalar(DstLV, CurCap->getLocation()); |
| 344 | } |
| 345 | CapturedVars.push_back(CV); |
| 346 | } else { |
| 347 | assert(CurCap->capturesVariable() && "Expected capture by reference." ); |
| 348 | CapturedVars.push_back(EmitLValue(*I).getAddress(*this).getPointer()); |
| 349 | } |
| 350 | } |
| 351 | } |
| 352 | |
| 353 | static Address castValueFromUintptr(CodeGenFunction &CGF, SourceLocation Loc, |
| 354 | QualType DstType, StringRef Name, |
| 355 | LValue AddrLV) { |
| 356 | ASTContext &Ctx = CGF.getContext(); |
| 357 | |
| 358 | llvm::Value *CastedPtr = CGF.EmitScalarConversion( |
| 359 | AddrLV.getAddress(CGF).getPointer(), Ctx.getUIntPtrType(), |
| 360 | Ctx.getPointerType(DstType), Loc); |
| 361 | Address TmpAddr = |
| 362 | CGF.MakeNaturalAlignAddrLValue(CastedPtr, Ctx.getPointerType(DstType)) |
| 363 | .getAddress(CGF); |
| 364 | return TmpAddr; |
| 365 | } |
| 366 | |
| 367 | static QualType getCanonicalParamType(ASTContext &C, QualType T) { |
| 368 | if (T->isLValueReferenceType()) |
| 369 | return C.getLValueReferenceType( |
| 370 | getCanonicalParamType(C, T.getNonReferenceType()), |
| 371 | /*SpelledAsLValue=*/false); |
| 372 | if (T->isPointerType()) |
| 373 | return C.getPointerType(getCanonicalParamType(C, T->getPointeeType())); |
| 374 | if (const ArrayType *A = T->getAsArrayTypeUnsafe()) { |
| 375 | if (const auto *VLA = dyn_cast<VariableArrayType>(A)) |
| 376 | return getCanonicalParamType(C, VLA->getElementType()); |
| 377 | if (!A->isVariablyModifiedType()) |
| 378 | return C.getCanonicalType(T); |
| 379 | } |
| 380 | return C.getCanonicalParamType(T); |
| 381 | } |
| 382 | |
| 383 | namespace { |
| 384 | /// Contains required data for proper outlined function codegen. |
| 385 | struct FunctionOptions { |
| 386 | /// Captured statement for which the function is generated. |
| 387 | const CapturedStmt *S = nullptr; |
| 388 | /// true if cast to/from UIntPtr is required for variables captured by |
| 389 | /// value. |
| 390 | const bool UIntPtrCastRequired = true; |
| 391 | /// true if only casted arguments must be registered as local args or VLA |
| 392 | /// sizes. |
| 393 | const bool RegisterCastedArgsOnly = false; |
| 394 | /// Name of the generated function. |
| 395 | const StringRef FunctionName; |
| 396 | /// Location of the non-debug version of the outlined function. |
| 397 | SourceLocation Loc; |
| 398 | explicit FunctionOptions(const CapturedStmt *S, bool UIntPtrCastRequired, |
| 399 | bool RegisterCastedArgsOnly, StringRef FunctionName, |
| 400 | SourceLocation Loc) |
| 401 | : S(S), UIntPtrCastRequired(UIntPtrCastRequired), |
| 402 | RegisterCastedArgsOnly(UIntPtrCastRequired && RegisterCastedArgsOnly), |
| 403 | FunctionName(FunctionName), Loc(Loc) {} |
| 404 | }; |
| 405 | } // namespace |
| 406 | |
| 407 | static llvm::Function *emitOutlinedFunctionPrologue( |
| 408 | CodeGenFunction &CGF, FunctionArgList &Args, |
| 409 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> |
| 410 | &LocalAddrs, |
| 411 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> |
| 412 | &VLASizes, |
| 413 | llvm::Value *&CXXThisValue, const FunctionOptions &FO) { |
| 414 | const CapturedDecl *CD = FO.S->getCapturedDecl(); |
| 415 | const RecordDecl *RD = FO.S->getCapturedRecordDecl(); |
| 416 | assert(CD->hasBody() && "missing CapturedDecl body" ); |
| 417 | |
| 418 | CXXThisValue = nullptr; |
| 419 | // Build the argument list. |
| 420 | CodeGenModule &CGM = CGF.CGM; |
| 421 | ASTContext &Ctx = CGM.getContext(); |
| 422 | FunctionArgList TargetArgs; |
| 423 | Args.append(CD->param_begin(), |
| 424 | std::next(CD->param_begin(), CD->getContextParamPosition())); |
| 425 | TargetArgs.append( |
| 426 | CD->param_begin(), |
| 427 | std::next(CD->param_begin(), CD->getContextParamPosition())); |
| 428 | auto I = FO.S->captures().begin(); |
| 429 | FunctionDecl *DebugFunctionDecl = nullptr; |
| 430 | if (!FO.UIntPtrCastRequired) { |
| 431 | FunctionProtoType::ExtProtoInfo EPI; |
| 432 | QualType FunctionTy = Ctx.getFunctionType(Ctx.VoidTy, llvm::None, EPI); |
| 433 | DebugFunctionDecl = FunctionDecl::Create( |
| 434 | Ctx, Ctx.getTranslationUnitDecl(), FO.S->getBeginLoc(), |
| 435 | SourceLocation(), DeclarationName(), FunctionTy, |
| 436 | Ctx.getTrivialTypeSourceInfo(FunctionTy), SC_Static, |
| 437 | /*isInlineSpecified=*/false, /*hasWrittenPrototype=*/false); |
| 438 | } |
| 439 | for (const FieldDecl *FD : RD->fields()) { |
| 440 | QualType ArgType = FD->getType(); |
| 441 | IdentifierInfo *II = nullptr; |
| 442 | VarDecl *CapVar = nullptr; |
| 443 | |
| 444 | // If this is a capture by copy and the type is not a pointer, the outlined |
| 445 | // function argument type should be uintptr and the value properly casted to |
| 446 | // uintptr. This is necessary given that the runtime library is only able to |
| 447 | // deal with pointers. We can pass in the same way the VLA type sizes to the |
| 448 | // outlined function. |
| 449 | if (FO.UIntPtrCastRequired && |
| 450 | ((I->capturesVariableByCopy() && !ArgType->isAnyPointerType()) || |
| 451 | I->capturesVariableArrayType())) |
| 452 | ArgType = Ctx.getUIntPtrType(); |
| 453 | |
| 454 | if (I->capturesVariable() || I->capturesVariableByCopy()) { |
| 455 | CapVar = I->getCapturedVar(); |
| 456 | II = CapVar->getIdentifier(); |
| 457 | } else if (I->capturesThis()) { |
| 458 | II = &Ctx.Idents.get("this" ); |
| 459 | } else { |
| 460 | assert(I->capturesVariableArrayType()); |
| 461 | II = &Ctx.Idents.get("vla" ); |
| 462 | } |
| 463 | if (ArgType->isVariablyModifiedType()) |
| 464 | ArgType = getCanonicalParamType(Ctx, ArgType); |
| 465 | VarDecl *Arg; |
| 466 | if (DebugFunctionDecl && (CapVar || I->capturesThis())) { |
| 467 | Arg = ParmVarDecl::Create( |
| 468 | Ctx, DebugFunctionDecl, |
| 469 | CapVar ? CapVar->getBeginLoc() : FD->getBeginLoc(), |
| 470 | CapVar ? CapVar->getLocation() : FD->getLocation(), II, ArgType, |
| 471 | /*TInfo=*/nullptr, SC_None, /*DefArg=*/nullptr); |
| 472 | } else { |
| 473 | Arg = ImplicitParamDecl::Create(Ctx, /*DC=*/nullptr, FD->getLocation(), |
| 474 | II, ArgType, ImplicitParamDecl::Other); |
| 475 | } |
| 476 | Args.emplace_back(Arg); |
| 477 | // Do not cast arguments if we emit function with non-original types. |
| 478 | TargetArgs.emplace_back( |
| 479 | FO.UIntPtrCastRequired |
| 480 | ? Arg |
| 481 | : CGM.getOpenMPRuntime().translateParameter(FD, Arg)); |
| 482 | ++I; |
| 483 | } |
| 484 | Args.append( |
| 485 | std::next(CD->param_begin(), CD->getContextParamPosition() + 1), |
| 486 | CD->param_end()); |
| 487 | TargetArgs.append( |
| 488 | std::next(CD->param_begin(), CD->getContextParamPosition() + 1), |
| 489 | CD->param_end()); |
| 490 | |
| 491 | // Create the function declaration. |
| 492 | const CGFunctionInfo &FuncInfo = |
| 493 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(Ctx.VoidTy, TargetArgs); |
| 494 | llvm::FunctionType *FuncLLVMTy = CGM.getTypes().GetFunctionType(FuncInfo); |
| 495 | |
| 496 | auto *F = |
| 497 | llvm::Function::Create(FuncLLVMTy, llvm::GlobalValue::InternalLinkage, |
| 498 | FO.FunctionName, &CGM.getModule()); |
| 499 | CGM.SetInternalFunctionAttributes(CD, F, FuncInfo); |
| 500 | if (CD->isNothrow()) |
| 501 | F->setDoesNotThrow(); |
| 502 | F->setDoesNotRecurse(); |
| 503 | |
| 504 | // Generate the function. |
| 505 | CGF.StartFunction(CD, Ctx.VoidTy, F, FuncInfo, TargetArgs, |
| 506 | FO.UIntPtrCastRequired ? FO.Loc : FO.S->getBeginLoc(), |
| 507 | FO.UIntPtrCastRequired ? FO.Loc |
| 508 | : CD->getBody()->getBeginLoc()); |
| 509 | unsigned Cnt = CD->getContextParamPosition(); |
| 510 | I = FO.S->captures().begin(); |
| 511 | for (const FieldDecl *FD : RD->fields()) { |
| 512 | // Do not map arguments if we emit function with non-original types. |
| 513 | Address LocalAddr(Address::invalid()); |
| 514 | if (!FO.UIntPtrCastRequired && Args[Cnt] != TargetArgs[Cnt]) { |
| 515 | LocalAddr = CGM.getOpenMPRuntime().getParameterAddress(CGF, Args[Cnt], |
| 516 | TargetArgs[Cnt]); |
| 517 | } else { |
| 518 | LocalAddr = CGF.GetAddrOfLocalVar(Args[Cnt]); |
| 519 | } |
| 520 | // If we are capturing a pointer by copy we don't need to do anything, just |
| 521 | // use the value that we get from the arguments. |
| 522 | if (I->capturesVariableByCopy() && FD->getType()->isAnyPointerType()) { |
| 523 | const VarDecl *CurVD = I->getCapturedVar(); |
| 524 | if (!FO.RegisterCastedArgsOnly) |
| 525 | LocalAddrs.insert({Args[Cnt], {CurVD, LocalAddr}}); |
| 526 | ++Cnt; |
| 527 | ++I; |
| 528 | continue; |
| 529 | } |
| 530 | |
| 531 | LValue ArgLVal = CGF.MakeAddrLValue(LocalAddr, Args[Cnt]->getType(), |
| 532 | AlignmentSource::Decl); |
| 533 | if (FD->hasCapturedVLAType()) { |
| 534 | if (FO.UIntPtrCastRequired) { |
| 535 | ArgLVal = CGF.MakeAddrLValue( |
| 536 | castValueFromUintptr(CGF, I->getLocation(), FD->getType(), |
| 537 | Args[Cnt]->getName(), ArgLVal), |
| 538 | FD->getType(), AlignmentSource::Decl); |
| 539 | } |
| 540 | llvm::Value *ExprArg = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation()); |
| 541 | const VariableArrayType *VAT = FD->getCapturedVLAType(); |
| 542 | VLASizes.try_emplace(Args[Cnt], VAT->getSizeExpr(), ExprArg); |
| 543 | } else if (I->capturesVariable()) { |
| 544 | const VarDecl *Var = I->getCapturedVar(); |
| 545 | QualType VarTy = Var->getType(); |
| 546 | Address ArgAddr = ArgLVal.getAddress(CGF); |
| 547 | if (ArgLVal.getType()->isLValueReferenceType()) { |
| 548 | ArgAddr = CGF.EmitLoadOfReference(ArgLVal); |
| 549 | } else if (!VarTy->isVariablyModifiedType() || !VarTy->isPointerType()) { |
| 550 | assert(ArgLVal.getType()->isPointerType()); |
| 551 | ArgAddr = CGF.EmitLoadOfPointer( |
| 552 | ArgAddr, ArgLVal.getType()->castAs<PointerType>()); |
| 553 | } |
| 554 | if (!FO.RegisterCastedArgsOnly) { |
| 555 | LocalAddrs.insert( |
| 556 | {Args[Cnt], |
| 557 | {Var, Address(ArgAddr.getPointer(), Ctx.getDeclAlign(Var))}}); |
| 558 | } |
| 559 | } else if (I->capturesVariableByCopy()) { |
| 560 | assert(!FD->getType()->isAnyPointerType() && |
| 561 | "Not expecting a captured pointer." ); |
| 562 | const VarDecl *Var = I->getCapturedVar(); |
| 563 | LocalAddrs.insert({Args[Cnt], |
| 564 | {Var, FO.UIntPtrCastRequired |
| 565 | ? castValueFromUintptr( |
| 566 | CGF, I->getLocation(), FD->getType(), |
| 567 | Args[Cnt]->getName(), ArgLVal) |
| 568 | : ArgLVal.getAddress(CGF)}}); |
| 569 | } else { |
| 570 | // If 'this' is captured, load it into CXXThisValue. |
| 571 | assert(I->capturesThis()); |
| 572 | CXXThisValue = CGF.EmitLoadOfScalar(ArgLVal, I->getLocation()); |
| 573 | LocalAddrs.insert({Args[Cnt], {nullptr, ArgLVal.getAddress(CGF)}}); |
| 574 | } |
| 575 | ++Cnt; |
| 576 | ++I; |
| 577 | } |
| 578 | |
| 579 | return F; |
| 580 | } |
| 581 | |
| 582 | llvm::Function * |
| 583 | CodeGenFunction::GenerateOpenMPCapturedStmtFunction(const CapturedStmt &S, |
| 584 | SourceLocation Loc) { |
| 585 | assert( |
| 586 | CapturedStmtInfo && |
| 587 | "CapturedStmtInfo should be set when generating the captured function" ); |
| 588 | const CapturedDecl *CD = S.getCapturedDecl(); |
| 589 | // Build the argument list. |
| 590 | bool NeedWrapperFunction = |
| 591 | getDebugInfo() && CGM.getCodeGenOpts().hasReducedDebugInfo(); |
| 592 | FunctionArgList Args; |
| 593 | llvm::MapVector<const Decl *, std::pair<const VarDecl *, Address>> LocalAddrs; |
| 594 | llvm::DenseMap<const Decl *, std::pair<const Expr *, llvm::Value *>> VLASizes; |
| 595 | SmallString<256> Buffer; |
| 596 | llvm::raw_svector_ostream Out(Buffer); |
| 597 | Out << CapturedStmtInfo->getHelperName(); |
| 598 | if (NeedWrapperFunction) |
| 599 | Out << "_debug__" ; |
| 600 | FunctionOptions FO(&S, !NeedWrapperFunction, /*RegisterCastedArgsOnly=*/false, |
| 601 | Out.str(), Loc); |
| 602 | llvm::Function *F = emitOutlinedFunctionPrologue(*this, Args, LocalAddrs, |
| 603 | VLASizes, CXXThisValue, FO); |
| 604 | CodeGenFunction::OMPPrivateScope LocalScope(*this); |
| 605 | for (const auto &LocalAddrPair : LocalAddrs) { |
| 606 | if (LocalAddrPair.second.first) { |
| 607 | LocalScope.addPrivate(LocalAddrPair.second.first, [&LocalAddrPair]() { |
| 608 | return LocalAddrPair.second.second; |
| 609 | }); |
| 610 | } |
| 611 | } |
| 612 | (void)LocalScope.Privatize(); |
| 613 | for (const auto &VLASizePair : VLASizes) |
| 614 | VLASizeMap[VLASizePair.second.first] = VLASizePair.second.second; |
| 615 | PGO.assignRegionCounters(GlobalDecl(CD), F); |
| 616 | CapturedStmtInfo->EmitBody(*this, CD->getBody()); |
| 617 | (void)LocalScope.ForceCleanup(); |
| 618 | FinishFunction(CD->getBodyRBrace()); |
| 619 | if (!NeedWrapperFunction) |
| 620 | return F; |
| 621 | |
| 622 | FunctionOptions WrapperFO(&S, /*UIntPtrCastRequired=*/true, |
| 623 | /*RegisterCastedArgsOnly=*/true, |
| 624 | CapturedStmtInfo->getHelperName(), Loc); |
| 625 | CodeGenFunction WrapperCGF(CGM, /*suppressNewContext=*/true); |
| 626 | WrapperCGF.CapturedStmtInfo = CapturedStmtInfo; |
| 627 | Args.clear(); |
| 628 | LocalAddrs.clear(); |
| 629 | VLASizes.clear(); |
| 630 | llvm::Function *WrapperF = |
| 631 | emitOutlinedFunctionPrologue(WrapperCGF, Args, LocalAddrs, VLASizes, |
| 632 | WrapperCGF.CXXThisValue, WrapperFO); |
| 633 | llvm::SmallVector<llvm::Value *, 4> CallArgs; |
| 634 | for (const auto *Arg : Args) { |
| 635 | llvm::Value *CallArg; |
| 636 | auto I = LocalAddrs.find(Arg); |
| 637 | if (I != LocalAddrs.end()) { |
| 638 | LValue LV = WrapperCGF.MakeAddrLValue( |
| 639 | I->second.second, |
| 640 | I->second.first ? I->second.first->getType() : Arg->getType(), |
| 641 | AlignmentSource::Decl); |
| 642 | CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc()); |
| 643 | } else { |
| 644 | auto EI = VLASizes.find(Arg); |
| 645 | if (EI != VLASizes.end()) { |
| 646 | CallArg = EI->second.second; |
| 647 | } else { |
| 648 | LValue LV = WrapperCGF.MakeAddrLValue(WrapperCGF.GetAddrOfLocalVar(Arg), |
| 649 | Arg->getType(), |
| 650 | AlignmentSource::Decl); |
| 651 | CallArg = WrapperCGF.EmitLoadOfScalar(LV, S.getBeginLoc()); |
| 652 | } |
| 653 | } |
| 654 | CallArgs.emplace_back(WrapperCGF.EmitFromMemory(CallArg, Arg->getType())); |
| 655 | } |
| 656 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(WrapperCGF, Loc, F, CallArgs); |
| 657 | WrapperCGF.FinishFunction(); |
| 658 | return WrapperF; |
| 659 | } |
| 660 | |
| 661 | //===----------------------------------------------------------------------===// |
| 662 | // OpenMP Directive Emission |
| 663 | //===----------------------------------------------------------------------===// |
| 664 | void CodeGenFunction::EmitOMPAggregateAssign( |
| 665 | Address DestAddr, Address SrcAddr, QualType OriginalType, |
| 666 | const llvm::function_ref<void(Address, Address)> CopyGen) { |
| 667 | // Perform element-by-element initialization. |
| 668 | QualType ElementTy; |
| 669 | |
| 670 | // Drill down to the base element type on both arrays. |
| 671 | const ArrayType *ArrayTy = OriginalType->getAsArrayTypeUnsafe(); |
| 672 | llvm::Value *NumElements = emitArrayLength(ArrayTy, ElementTy, DestAddr); |
| 673 | SrcAddr = Builder.CreateElementBitCast(SrcAddr, DestAddr.getElementType()); |
| 674 | |
| 675 | llvm::Value *SrcBegin = SrcAddr.getPointer(); |
| 676 | llvm::Value *DestBegin = DestAddr.getPointer(); |
| 677 | // Cast from pointer to array type to pointer to single element. |
| 678 | llvm::Value *DestEnd = Builder.CreateGEP(DestBegin, NumElements); |
| 679 | // The basic structure here is a while-do loop. |
| 680 | llvm::BasicBlock *BodyBB = createBasicBlock("omp.arraycpy.body" ); |
| 681 | llvm::BasicBlock *DoneBB = createBasicBlock("omp.arraycpy.done" ); |
| 682 | llvm::Value *IsEmpty = |
| 683 | Builder.CreateICmpEQ(DestBegin, DestEnd, "omp.arraycpy.isempty" ); |
| 684 | Builder.CreateCondBr(IsEmpty, DoneBB, BodyBB); |
| 685 | |
| 686 | // Enter the loop body, making that address the current address. |
| 687 | llvm::BasicBlock *EntryBB = Builder.GetInsertBlock(); |
| 688 | EmitBlock(BodyBB); |
| 689 | |
| 690 | CharUnits ElementSize = getContext().getTypeSizeInChars(ElementTy); |
| 691 | |
| 692 | llvm::PHINode *SrcElementPHI = |
| 693 | Builder.CreatePHI(SrcBegin->getType(), 2, "omp.arraycpy.srcElementPast" ); |
| 694 | SrcElementPHI->addIncoming(SrcBegin, EntryBB); |
| 695 | Address SrcElementCurrent = |
| 696 | Address(SrcElementPHI, |
| 697 | SrcAddr.getAlignment().alignmentOfArrayElement(ElementSize)); |
| 698 | |
| 699 | llvm::PHINode *DestElementPHI = |
| 700 | Builder.CreatePHI(DestBegin->getType(), 2, "omp.arraycpy.destElementPast" ); |
| 701 | DestElementPHI->addIncoming(DestBegin, EntryBB); |
| 702 | Address DestElementCurrent = |
| 703 | Address(DestElementPHI, |
| 704 | DestAddr.getAlignment().alignmentOfArrayElement(ElementSize)); |
| 705 | |
| 706 | // Emit copy. |
| 707 | CopyGen(DestElementCurrent, SrcElementCurrent); |
| 708 | |
| 709 | // Shift the address forward by one element. |
| 710 | llvm::Value *DestElementNext = Builder.CreateConstGEP1_32( |
| 711 | DestElementPHI, /*Idx0=*/1, "omp.arraycpy.dest.element" ); |
| 712 | llvm::Value *SrcElementNext = Builder.CreateConstGEP1_32( |
| 713 | SrcElementPHI, /*Idx0=*/1, "omp.arraycpy.src.element" ); |
| 714 | // Check whether we've reached the end. |
| 715 | llvm::Value *Done = |
| 716 | Builder.CreateICmpEQ(DestElementNext, DestEnd, "omp.arraycpy.done" ); |
| 717 | Builder.CreateCondBr(Done, DoneBB, BodyBB); |
| 718 | DestElementPHI->addIncoming(DestElementNext, Builder.GetInsertBlock()); |
| 719 | SrcElementPHI->addIncoming(SrcElementNext, Builder.GetInsertBlock()); |
| 720 | |
| 721 | // Done. |
| 722 | EmitBlock(DoneBB, /*IsFinished=*/true); |
| 723 | } |
| 724 | |
| 725 | void CodeGenFunction::EmitOMPCopy(QualType OriginalType, Address DestAddr, |
| 726 | Address SrcAddr, const VarDecl *DestVD, |
| 727 | const VarDecl *SrcVD, const Expr *Copy) { |
| 728 | if (OriginalType->isArrayType()) { |
| 729 | const auto *BO = dyn_cast<BinaryOperator>(Copy); |
| 730 | if (BO && BO->getOpcode() == BO_Assign) { |
| 731 | // Perform simple memcpy for simple copying. |
| 732 | LValue Dest = MakeAddrLValue(DestAddr, OriginalType); |
| 733 | LValue Src = MakeAddrLValue(SrcAddr, OriginalType); |
| 734 | EmitAggregateAssign(Dest, Src, OriginalType); |
| 735 | } else { |
| 736 | // For arrays with complex element types perform element by element |
| 737 | // copying. |
| 738 | EmitOMPAggregateAssign( |
| 739 | DestAddr, SrcAddr, OriginalType, |
| 740 | [this, Copy, SrcVD, DestVD](Address DestElement, Address SrcElement) { |
| 741 | // Working with the single array element, so have to remap |
| 742 | // destination and source variables to corresponding array |
| 743 | // elements. |
| 744 | CodeGenFunction::OMPPrivateScope Remap(*this); |
| 745 | Remap.addPrivate(DestVD, [DestElement]() { return DestElement; }); |
| 746 | Remap.addPrivate(SrcVD, [SrcElement]() { return SrcElement; }); |
| 747 | (void)Remap.Privatize(); |
| 748 | EmitIgnoredExpr(Copy); |
| 749 | }); |
| 750 | } |
| 751 | } else { |
| 752 | // Remap pseudo source variable to private copy. |
| 753 | CodeGenFunction::OMPPrivateScope Remap(*this); |
| 754 | Remap.addPrivate(SrcVD, [SrcAddr]() { return SrcAddr; }); |
| 755 | Remap.addPrivate(DestVD, [DestAddr]() { return DestAddr; }); |
| 756 | (void)Remap.Privatize(); |
| 757 | // Emit copying of the whole variable. |
| 758 | EmitIgnoredExpr(Copy); |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | bool CodeGenFunction::EmitOMPFirstprivateClause(const OMPExecutableDirective &D, |
| 763 | OMPPrivateScope &PrivateScope) { |
| 764 | if (!HaveInsertPoint()) |
| 765 | return false; |
| 766 | bool DeviceConstTarget = |
| 767 | getLangOpts().OpenMPIsDevice && |
| 768 | isOpenMPTargetExecutionDirective(D.getDirectiveKind()); |
| 769 | bool FirstprivateIsLastprivate = false; |
| 770 | llvm::DenseMap<const VarDecl *, OpenMPLastprivateModifier> Lastprivates; |
| 771 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 772 | for (const auto *D : C->varlists()) |
| 773 | Lastprivates.try_emplace( |
| 774 | cast<VarDecl>(cast<DeclRefExpr>(D)->getDecl())->getCanonicalDecl(), |
| 775 | C->getKind()); |
| 776 | } |
| 777 | llvm::DenseSet<const VarDecl *> EmittedAsFirstprivate; |
| 778 | llvm::SmallVector<OpenMPDirectiveKind, 4> CaptureRegions; |
| 779 | getOpenMPCaptureRegions(CaptureRegions, D.getDirectiveKind()); |
| 780 | // Force emission of the firstprivate copy if the directive does not emit |
| 781 | // outlined function, like omp for, omp simd, omp distribute etc. |
| 782 | bool MustEmitFirstprivateCopy = |
| 783 | CaptureRegions.size() == 1 && CaptureRegions.back() == OMPD_unknown; |
| 784 | for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) { |
| 785 | const auto *IRef = C->varlist_begin(); |
| 786 | const auto *InitsRef = C->inits().begin(); |
| 787 | for (const Expr *IInit : C->private_copies()) { |
| 788 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 789 | bool ThisFirstprivateIsLastprivate = |
| 790 | Lastprivates.count(OrigVD->getCanonicalDecl()) > 0; |
| 791 | const FieldDecl *FD = CapturedStmtInfo->lookup(OrigVD); |
| 792 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl()); |
| 793 | if (!MustEmitFirstprivateCopy && !ThisFirstprivateIsLastprivate && FD && |
| 794 | !FD->getType()->isReferenceType() && |
| 795 | (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) { |
| 796 | EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl()); |
| 797 | ++IRef; |
| 798 | ++InitsRef; |
| 799 | continue; |
| 800 | } |
| 801 | // Do not emit copy for firstprivate constant variables in target regions, |
| 802 | // captured by reference. |
| 803 | if (DeviceConstTarget && OrigVD->getType().isConstant(getContext()) && |
| 804 | FD && FD->getType()->isReferenceType() && |
| 805 | (!VD || !VD->hasAttr<OMPAllocateDeclAttr>())) { |
| 806 | (void)CGM.getOpenMPRuntime().registerTargetFirstprivateCopy(*this, |
| 807 | OrigVD); |
| 808 | ++IRef; |
| 809 | ++InitsRef; |
| 810 | continue; |
| 811 | } |
| 812 | FirstprivateIsLastprivate = |
| 813 | FirstprivateIsLastprivate || ThisFirstprivateIsLastprivate; |
| 814 | if (EmittedAsFirstprivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 815 | const auto *VDInit = |
| 816 | cast<VarDecl>(cast<DeclRefExpr>(*InitsRef)->getDecl()); |
| 817 | bool IsRegistered; |
| 818 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 819 | /*RefersToEnclosingVariableOrCapture=*/FD != nullptr, |
| 820 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 821 | LValue OriginalLVal; |
| 822 | if (!FD) { |
| 823 | // Check if the firstprivate variable is just a constant value. |
| 824 | ConstantEmission CE = tryEmitAsConstant(&DRE); |
| 825 | if (CE && !CE.isReference()) { |
| 826 | // Constant value, no need to create a copy. |
| 827 | ++IRef; |
| 828 | ++InitsRef; |
| 829 | continue; |
| 830 | } |
| 831 | if (CE && CE.isReference()) { |
| 832 | OriginalLVal = CE.getReferenceLValue(*this, &DRE); |
| 833 | } else { |
| 834 | assert(!CE && "Expected non-constant firstprivate." ); |
| 835 | OriginalLVal = EmitLValue(&DRE); |
| 836 | } |
| 837 | } else { |
| 838 | OriginalLVal = EmitLValue(&DRE); |
| 839 | } |
| 840 | QualType Type = VD->getType(); |
| 841 | if (Type->isArrayType()) { |
| 842 | // Emit VarDecl with copy init for arrays. |
| 843 | // Get the address of the original variable captured in current |
| 844 | // captured region. |
| 845 | IsRegistered = PrivateScope.addPrivate( |
| 846 | OrigVD, [this, VD, Type, OriginalLVal, VDInit]() { |
| 847 | AutoVarEmission Emission = EmitAutoVarAlloca(*VD); |
| 848 | const Expr *Init = VD->getInit(); |
| 849 | if (!isa<CXXConstructExpr>(Init) || |
| 850 | isTrivialInitializer(Init)) { |
| 851 | // Perform simple memcpy. |
| 852 | LValue Dest = |
| 853 | MakeAddrLValue(Emission.getAllocatedAddress(), Type); |
| 854 | EmitAggregateAssign(Dest, OriginalLVal, Type); |
| 855 | } else { |
| 856 | EmitOMPAggregateAssign( |
| 857 | Emission.getAllocatedAddress(), |
| 858 | OriginalLVal.getAddress(*this), Type, |
| 859 | [this, VDInit, Init](Address DestElement, |
| 860 | Address SrcElement) { |
| 861 | // Clean up any temporaries needed by the |
| 862 | // initialization. |
| 863 | RunCleanupsScope InitScope(*this); |
| 864 | // Emit initialization for single element. |
| 865 | setAddrOfLocalVar(VDInit, SrcElement); |
| 866 | EmitAnyExprToMem(Init, DestElement, |
| 867 | Init->getType().getQualifiers(), |
| 868 | /*IsInitializer*/ false); |
| 869 | LocalDeclMap.erase(VDInit); |
| 870 | }); |
| 871 | } |
| 872 | EmitAutoVarCleanups(Emission); |
| 873 | return Emission.getAllocatedAddress(); |
| 874 | }); |
| 875 | } else { |
| 876 | Address OriginalAddr = OriginalLVal.getAddress(*this); |
| 877 | IsRegistered = |
| 878 | PrivateScope.addPrivate(OrigVD, [this, VDInit, OriginalAddr, VD, |
| 879 | ThisFirstprivateIsLastprivate, |
| 880 | OrigVD, &Lastprivates, IRef]() { |
| 881 | // Emit private VarDecl with copy init. |
| 882 | // Remap temp VDInit variable to the address of the original |
| 883 | // variable (for proper handling of captured global variables). |
| 884 | setAddrOfLocalVar(VDInit, OriginalAddr); |
| 885 | EmitDecl(*VD); |
| 886 | LocalDeclMap.erase(VDInit); |
| 887 | if (ThisFirstprivateIsLastprivate && |
| 888 | Lastprivates[OrigVD->getCanonicalDecl()] == |
| 889 | OMPC_LASTPRIVATE_conditional) { |
| 890 | // Create/init special variable for lastprivate conditionals. |
| 891 | Address VDAddr = |
| 892 | CGM.getOpenMPRuntime().emitLastprivateConditionalInit( |
| 893 | *this, OrigVD); |
| 894 | llvm::Value *V = EmitLoadOfScalar( |
| 895 | MakeAddrLValue(GetAddrOfLocalVar(VD), (*IRef)->getType(), |
| 896 | AlignmentSource::Decl), |
| 897 | (*IRef)->getExprLoc()); |
| 898 | EmitStoreOfScalar(V, |
| 899 | MakeAddrLValue(VDAddr, (*IRef)->getType(), |
| 900 | AlignmentSource::Decl)); |
| 901 | LocalDeclMap.erase(VD); |
| 902 | setAddrOfLocalVar(VD, VDAddr); |
| 903 | return VDAddr; |
| 904 | } |
| 905 | return GetAddrOfLocalVar(VD); |
| 906 | }); |
| 907 | } |
| 908 | assert(IsRegistered && |
| 909 | "firstprivate var already registered as private" ); |
| 910 | // Silence the warning about unused variable. |
| 911 | (void)IsRegistered; |
| 912 | } |
| 913 | ++IRef; |
| 914 | ++InitsRef; |
| 915 | } |
| 916 | } |
| 917 | return FirstprivateIsLastprivate && !EmittedAsFirstprivate.empty(); |
| 918 | } |
| 919 | |
| 920 | void CodeGenFunction::EmitOMPPrivateClause( |
| 921 | const OMPExecutableDirective &D, |
| 922 | CodeGenFunction::OMPPrivateScope &PrivateScope) { |
| 923 | if (!HaveInsertPoint()) |
| 924 | return; |
| 925 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 926 | for (const auto *C : D.getClausesOfKind<OMPPrivateClause>()) { |
| 927 | auto IRef = C->varlist_begin(); |
| 928 | for (const Expr *IInit : C->private_copies()) { |
| 929 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 930 | if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 931 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl()); |
| 932 | bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD]() { |
| 933 | // Emit private VarDecl with copy init. |
| 934 | EmitDecl(*VD); |
| 935 | return GetAddrOfLocalVar(VD); |
| 936 | }); |
| 937 | assert(IsRegistered && "private var already registered as private" ); |
| 938 | // Silence the warning about unused variable. |
| 939 | (void)IsRegistered; |
| 940 | } |
| 941 | ++IRef; |
| 942 | } |
| 943 | } |
| 944 | } |
| 945 | |
| 946 | bool CodeGenFunction::EmitOMPCopyinClause(const OMPExecutableDirective &D) { |
| 947 | if (!HaveInsertPoint()) |
| 948 | return false; |
| 949 | // threadprivate_var1 = master_threadprivate_var1; |
| 950 | // operator=(threadprivate_var2, master_threadprivate_var2); |
| 951 | // ... |
| 952 | // __kmpc_barrier(&loc, global_tid); |
| 953 | llvm::DenseSet<const VarDecl *> CopiedVars; |
| 954 | llvm::BasicBlock *CopyBegin = nullptr, *CopyEnd = nullptr; |
| 955 | for (const auto *C : D.getClausesOfKind<OMPCopyinClause>()) { |
| 956 | auto IRef = C->varlist_begin(); |
| 957 | auto ISrcRef = C->source_exprs().begin(); |
| 958 | auto IDestRef = C->destination_exprs().begin(); |
| 959 | for (const Expr *AssignOp : C->assignment_ops()) { |
| 960 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 961 | QualType Type = VD->getType(); |
| 962 | if (CopiedVars.insert(VD->getCanonicalDecl()).second) { |
| 963 | // Get the address of the master variable. If we are emitting code with |
| 964 | // TLS support, the address is passed from the master as field in the |
| 965 | // captured declaration. |
| 966 | Address MasterAddr = Address::invalid(); |
| 967 | if (getLangOpts().OpenMPUseTLS && |
| 968 | getContext().getTargetInfo().isTLSSupported()) { |
| 969 | assert(CapturedStmtInfo->lookup(VD) && |
| 970 | "Copyin threadprivates should have been captured!" ); |
| 971 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), true, |
| 972 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 973 | MasterAddr = EmitLValue(&DRE).getAddress(*this); |
| 974 | LocalDeclMap.erase(VD); |
| 975 | } else { |
| 976 | MasterAddr = |
| 977 | Address(VD->isStaticLocal() ? CGM.getStaticLocalDeclAddress(VD) |
| 978 | : CGM.GetAddrOfGlobal(VD), |
| 979 | getContext().getDeclAlign(VD)); |
| 980 | } |
| 981 | // Get the address of the threadprivate variable. |
| 982 | Address PrivateAddr = EmitLValue(*IRef).getAddress(*this); |
| 983 | if (CopiedVars.size() == 1) { |
| 984 | // At first check if current thread is a master thread. If it is, no |
| 985 | // need to copy data. |
| 986 | CopyBegin = createBasicBlock("copyin.not.master" ); |
| 987 | CopyEnd = createBasicBlock("copyin.not.master.end" ); |
| 988 | Builder.CreateCondBr( |
| 989 | Builder.CreateICmpNE( |
| 990 | Builder.CreatePtrToInt(MasterAddr.getPointer(), CGM.IntPtrTy), |
| 991 | Builder.CreatePtrToInt(PrivateAddr.getPointer(), |
| 992 | CGM.IntPtrTy)), |
| 993 | CopyBegin, CopyEnd); |
| 994 | EmitBlock(CopyBegin); |
| 995 | } |
| 996 | const auto *SrcVD = |
| 997 | cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl()); |
| 998 | const auto *DestVD = |
| 999 | cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl()); |
| 1000 | EmitOMPCopy(Type, PrivateAddr, MasterAddr, DestVD, SrcVD, AssignOp); |
| 1001 | } |
| 1002 | ++IRef; |
| 1003 | ++ISrcRef; |
| 1004 | ++IDestRef; |
| 1005 | } |
| 1006 | } |
| 1007 | if (CopyEnd) { |
| 1008 | // Exit out of copying procedure for non-master thread. |
| 1009 | EmitBlock(CopyEnd, /*IsFinished=*/true); |
| 1010 | return true; |
| 1011 | } |
| 1012 | return false; |
| 1013 | } |
| 1014 | |
| 1015 | bool CodeGenFunction::EmitOMPLastprivateClauseInit( |
| 1016 | const OMPExecutableDirective &D, OMPPrivateScope &PrivateScope) { |
| 1017 | if (!HaveInsertPoint()) |
| 1018 | return false; |
| 1019 | bool HasAtLeastOneLastprivate = false; |
| 1020 | llvm::DenseSet<const VarDecl *> SIMDLCVs; |
| 1021 | if (isOpenMPSimdDirective(D.getDirectiveKind())) { |
| 1022 | const auto *LoopDirective = cast<OMPLoopDirective>(&D); |
| 1023 | for (const Expr *C : LoopDirective->counters()) { |
| 1024 | SIMDLCVs.insert( |
| 1025 | cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl()); |
| 1026 | } |
| 1027 | } |
| 1028 | llvm::DenseSet<const VarDecl *> AlreadyEmittedVars; |
| 1029 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 1030 | HasAtLeastOneLastprivate = true; |
| 1031 | if (isOpenMPTaskLoopDirective(D.getDirectiveKind()) && |
| 1032 | !getLangOpts().OpenMPSimd) |
| 1033 | break; |
| 1034 | const auto *IRef = C->varlist_begin(); |
| 1035 | const auto *IDestRef = C->destination_exprs().begin(); |
| 1036 | for (const Expr *IInit : C->private_copies()) { |
| 1037 | // Keep the address of the original variable for future update at the end |
| 1038 | // of the loop. |
| 1039 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 1040 | // Taskloops do not require additional initialization, it is done in |
| 1041 | // runtime support library. |
| 1042 | if (AlreadyEmittedVars.insert(OrigVD->getCanonicalDecl()).second) { |
| 1043 | const auto *DestVD = |
| 1044 | cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl()); |
| 1045 | PrivateScope.addPrivate(DestVD, [this, OrigVD, IRef]() { |
| 1046 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 1047 | /*RefersToEnclosingVariableOrCapture=*/ |
| 1048 | CapturedStmtInfo->lookup(OrigVD) != nullptr, |
| 1049 | (*IRef)->getType(), VK_LValue, (*IRef)->getExprLoc()); |
| 1050 | return EmitLValue(&DRE).getAddress(*this); |
| 1051 | }); |
| 1052 | // Check if the variable is also a firstprivate: in this case IInit is |
| 1053 | // not generated. Initialization of this variable will happen in codegen |
| 1054 | // for 'firstprivate' clause. |
| 1055 | if (IInit && !SIMDLCVs.count(OrigVD->getCanonicalDecl())) { |
| 1056 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(IInit)->getDecl()); |
| 1057 | bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, VD, C, |
| 1058 | OrigVD]() { |
| 1059 | if (C->getKind() == OMPC_LASTPRIVATE_conditional) { |
| 1060 | Address VDAddr = |
| 1061 | CGM.getOpenMPRuntime().emitLastprivateConditionalInit(*this, |
| 1062 | OrigVD); |
| 1063 | setAddrOfLocalVar(VD, VDAddr); |
| 1064 | return VDAddr; |
| 1065 | } |
| 1066 | // Emit private VarDecl with copy init. |
| 1067 | EmitDecl(*VD); |
| 1068 | return GetAddrOfLocalVar(VD); |
| 1069 | }); |
| 1070 | assert(IsRegistered && |
| 1071 | "lastprivate var already registered as private" ); |
| 1072 | (void)IsRegistered; |
| 1073 | } |
| 1074 | } |
| 1075 | ++IRef; |
| 1076 | ++IDestRef; |
| 1077 | } |
| 1078 | } |
| 1079 | return HasAtLeastOneLastprivate; |
| 1080 | } |
| 1081 | |
| 1082 | void CodeGenFunction::EmitOMPLastprivateClauseFinal( |
| 1083 | const OMPExecutableDirective &D, bool NoFinals, |
| 1084 | llvm::Value *IsLastIterCond) { |
| 1085 | if (!HaveInsertPoint()) |
| 1086 | return; |
| 1087 | // Emit following code: |
| 1088 | // if (<IsLastIterCond>) { |
| 1089 | // orig_var1 = private_orig_var1; |
| 1090 | // ... |
| 1091 | // orig_varn = private_orig_varn; |
| 1092 | // } |
| 1093 | llvm::BasicBlock *ThenBB = nullptr; |
| 1094 | llvm::BasicBlock *DoneBB = nullptr; |
| 1095 | if (IsLastIterCond) { |
| 1096 | // Emit implicit barrier if at least one lastprivate conditional is found |
| 1097 | // and this is not a simd mode. |
| 1098 | if (!getLangOpts().OpenMPSimd && |
| 1099 | llvm::any_of(D.getClausesOfKind<OMPLastprivateClause>(), |
| 1100 | [](const OMPLastprivateClause *C) { |
| 1101 | return C->getKind() == OMPC_LASTPRIVATE_conditional; |
| 1102 | })) { |
| 1103 | CGM.getOpenMPRuntime().emitBarrierCall(*this, D.getBeginLoc(), |
| 1104 | OMPD_unknown, |
| 1105 | /*EmitChecks=*/false, |
| 1106 | /*ForceSimpleCall=*/true); |
| 1107 | } |
| 1108 | ThenBB = createBasicBlock(".omp.lastprivate.then" ); |
| 1109 | DoneBB = createBasicBlock(".omp.lastprivate.done" ); |
| 1110 | Builder.CreateCondBr(IsLastIterCond, ThenBB, DoneBB); |
| 1111 | EmitBlock(ThenBB); |
| 1112 | } |
| 1113 | llvm::DenseSet<const VarDecl *> AlreadyEmittedVars; |
| 1114 | llvm::DenseMap<const VarDecl *, const Expr *> LoopCountersAndUpdates; |
| 1115 | if (const auto *LoopDirective = dyn_cast<OMPLoopDirective>(&D)) { |
| 1116 | auto IC = LoopDirective->counters().begin(); |
| 1117 | for (const Expr *F : LoopDirective->finals()) { |
| 1118 | const auto *D = |
| 1119 | cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl())->getCanonicalDecl(); |
| 1120 | if (NoFinals) |
| 1121 | AlreadyEmittedVars.insert(D); |
| 1122 | else |
| 1123 | LoopCountersAndUpdates[D] = F; |
| 1124 | ++IC; |
| 1125 | } |
| 1126 | } |
| 1127 | for (const auto *C : D.getClausesOfKind<OMPLastprivateClause>()) { |
| 1128 | auto IRef = C->varlist_begin(); |
| 1129 | auto ISrcRef = C->source_exprs().begin(); |
| 1130 | auto IDestRef = C->destination_exprs().begin(); |
| 1131 | for (const Expr *AssignOp : C->assignment_ops()) { |
| 1132 | const auto *PrivateVD = |
| 1133 | cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 1134 | QualType Type = PrivateVD->getType(); |
| 1135 | const auto *CanonicalVD = PrivateVD->getCanonicalDecl(); |
| 1136 | if (AlreadyEmittedVars.insert(CanonicalVD).second) { |
| 1137 | // If lastprivate variable is a loop control variable for loop-based |
| 1138 | // directive, update its value before copyin back to original |
| 1139 | // variable. |
| 1140 | if (const Expr *FinalExpr = LoopCountersAndUpdates.lookup(CanonicalVD)) |
| 1141 | EmitIgnoredExpr(FinalExpr); |
| 1142 | const auto *SrcVD = |
| 1143 | cast<VarDecl>(cast<DeclRefExpr>(*ISrcRef)->getDecl()); |
| 1144 | const auto *DestVD = |
| 1145 | cast<VarDecl>(cast<DeclRefExpr>(*IDestRef)->getDecl()); |
| 1146 | // Get the address of the private variable. |
| 1147 | Address PrivateAddr = GetAddrOfLocalVar(PrivateVD); |
| 1148 | if (const auto *RefTy = PrivateVD->getType()->getAs<ReferenceType>()) |
| 1149 | PrivateAddr = |
| 1150 | Address(Builder.CreateLoad(PrivateAddr), |
| 1151 | CGM.getNaturalTypeAlignment(RefTy->getPointeeType())); |
| 1152 | // Store the last value to the private copy in the last iteration. |
| 1153 | if (C->getKind() == OMPC_LASTPRIVATE_conditional) |
| 1154 | CGM.getOpenMPRuntime().emitLastprivateConditionalFinalUpdate( |
| 1155 | *this, MakeAddrLValue(PrivateAddr, (*IRef)->getType()), PrivateVD, |
| 1156 | (*IRef)->getExprLoc()); |
| 1157 | // Get the address of the original variable. |
| 1158 | Address OriginalAddr = GetAddrOfLocalVar(DestVD); |
| 1159 | EmitOMPCopy(Type, OriginalAddr, PrivateAddr, DestVD, SrcVD, AssignOp); |
| 1160 | } |
| 1161 | ++IRef; |
| 1162 | ++ISrcRef; |
| 1163 | ++IDestRef; |
| 1164 | } |
| 1165 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) |
| 1166 | EmitIgnoredExpr(PostUpdate); |
| 1167 | } |
| 1168 | if (IsLastIterCond) |
| 1169 | EmitBlock(DoneBB, /*IsFinished=*/true); |
| 1170 | } |
| 1171 | |
| 1172 | void CodeGenFunction::EmitOMPReductionClauseInit( |
| 1173 | const OMPExecutableDirective &D, |
| 1174 | CodeGenFunction::OMPPrivateScope &PrivateScope, bool ForInscan) { |
| 1175 | if (!HaveInsertPoint()) |
| 1176 | return; |
| 1177 | SmallVector<const Expr *, 4> Shareds; |
| 1178 | SmallVector<const Expr *, 4> Privates; |
| 1179 | SmallVector<const Expr *, 4> ReductionOps; |
| 1180 | SmallVector<const Expr *, 4> LHSs; |
| 1181 | SmallVector<const Expr *, 4> RHSs; |
| 1182 | OMPTaskDataTy Data; |
| 1183 | SmallVector<const Expr *, 4> TaskLHSs; |
| 1184 | SmallVector<const Expr *, 4> TaskRHSs; |
| 1185 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1186 | if (ForInscan != (C->getModifier() == OMPC_REDUCTION_inscan)) |
| 1187 | continue; |
| 1188 | Shareds.append(C->varlist_begin(), C->varlist_end()); |
| 1189 | Privates.append(C->privates().begin(), C->privates().end()); |
| 1190 | ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end()); |
| 1191 | LHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 1192 | RHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 1193 | if (C->getModifier() == OMPC_REDUCTION_task) { |
| 1194 | Data.ReductionVars.append(C->privates().begin(), C->privates().end()); |
| 1195 | Data.ReductionOrigs.append(C->varlist_begin(), C->varlist_end()); |
| 1196 | Data.ReductionCopies.append(C->privates().begin(), C->privates().end()); |
| 1197 | Data.ReductionOps.append(C->reduction_ops().begin(), |
| 1198 | C->reduction_ops().end()); |
| 1199 | TaskLHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 1200 | TaskRHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 1201 | } |
| 1202 | } |
| 1203 | ReductionCodeGen RedCG(Shareds, Shareds, Privates, ReductionOps); |
| 1204 | unsigned Count = 0; |
| 1205 | auto *ILHS = LHSs.begin(); |
| 1206 | auto *IRHS = RHSs.begin(); |
| 1207 | auto *IPriv = Privates.begin(); |
| 1208 | for (const Expr *IRef : Shareds) { |
| 1209 | const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*IPriv)->getDecl()); |
| 1210 | // Emit private VarDecl with reduction init. |
| 1211 | RedCG.emitSharedOrigLValue(*this, Count); |
| 1212 | RedCG.emitAggregateType(*this, Count); |
| 1213 | AutoVarEmission Emission = EmitAutoVarAlloca(*PrivateVD); |
| 1214 | RedCG.emitInitialization(*this, Count, Emission.getAllocatedAddress(), |
| 1215 | RedCG.getSharedLValue(Count), |
| 1216 | [&Emission](CodeGenFunction &CGF) { |
| 1217 | CGF.EmitAutoVarInit(Emission); |
| 1218 | return true; |
| 1219 | }); |
| 1220 | EmitAutoVarCleanups(Emission); |
| 1221 | Address BaseAddr = RedCG.adjustPrivateAddress( |
| 1222 | *this, Count, Emission.getAllocatedAddress()); |
| 1223 | bool IsRegistered = PrivateScope.addPrivate( |
| 1224 | RedCG.getBaseDecl(Count), [BaseAddr]() { return BaseAddr; }); |
| 1225 | assert(IsRegistered && "private var already registered as private" ); |
| 1226 | // Silence the warning about unused variable. |
| 1227 | (void)IsRegistered; |
| 1228 | |
| 1229 | const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); |
| 1230 | const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); |
| 1231 | QualType Type = PrivateVD->getType(); |
| 1232 | bool isaOMPArraySectionExpr = isa<OMPArraySectionExpr>(IRef); |
| 1233 | if (isaOMPArraySectionExpr && Type->isVariablyModifiedType()) { |
| 1234 | // Store the address of the original variable associated with the LHS |
| 1235 | // implicit variable. |
| 1236 | PrivateScope.addPrivate(LHSVD, [&RedCG, Count, this]() { |
| 1237 | return RedCG.getSharedLValue(Count).getAddress(*this); |
| 1238 | }); |
| 1239 | PrivateScope.addPrivate( |
| 1240 | RHSVD, [this, PrivateVD]() { return GetAddrOfLocalVar(PrivateVD); }); |
| 1241 | } else if ((isaOMPArraySectionExpr && Type->isScalarType()) || |
| 1242 | isa<ArraySubscriptExpr>(IRef)) { |
| 1243 | // Store the address of the original variable associated with the LHS |
| 1244 | // implicit variable. |
| 1245 | PrivateScope.addPrivate(LHSVD, [&RedCG, Count, this]() { |
| 1246 | return RedCG.getSharedLValue(Count).getAddress(*this); |
| 1247 | }); |
| 1248 | PrivateScope.addPrivate(RHSVD, [this, PrivateVD, RHSVD]() { |
| 1249 | return Builder.CreateElementBitCast(GetAddrOfLocalVar(PrivateVD), |
| 1250 | ConvertTypeForMem(RHSVD->getType()), |
| 1251 | "rhs.begin" ); |
| 1252 | }); |
| 1253 | } else { |
| 1254 | QualType Type = PrivateVD->getType(); |
| 1255 | bool IsArray = getContext().getAsArrayType(Type) != nullptr; |
| 1256 | Address OriginalAddr = RedCG.getSharedLValue(Count).getAddress(*this); |
| 1257 | // Store the address of the original variable associated with the LHS |
| 1258 | // implicit variable. |
| 1259 | if (IsArray) { |
| 1260 | OriginalAddr = Builder.CreateElementBitCast( |
| 1261 | OriginalAddr, ConvertTypeForMem(LHSVD->getType()), "lhs.begin" ); |
| 1262 | } |
| 1263 | PrivateScope.addPrivate(LHSVD, [OriginalAddr]() { return OriginalAddr; }); |
| 1264 | PrivateScope.addPrivate( |
| 1265 | RHSVD, [this, PrivateVD, RHSVD, IsArray]() { |
| 1266 | return IsArray |
| 1267 | ? Builder.CreateElementBitCast( |
| 1268 | GetAddrOfLocalVar(PrivateVD), |
| 1269 | ConvertTypeForMem(RHSVD->getType()), "rhs.begin" ) |
| 1270 | : GetAddrOfLocalVar(PrivateVD); |
| 1271 | }); |
| 1272 | } |
| 1273 | ++ILHS; |
| 1274 | ++IRHS; |
| 1275 | ++IPriv; |
| 1276 | ++Count; |
| 1277 | } |
| 1278 | if (!Data.ReductionVars.empty()) { |
| 1279 | Data.IsReductionWithTaskMod = true; |
| 1280 | Data.IsWorksharingReduction = |
| 1281 | isOpenMPWorksharingDirective(D.getDirectiveKind()); |
| 1282 | llvm::Value *ReductionDesc = CGM.getOpenMPRuntime().emitTaskReductionInit( |
| 1283 | *this, D.getBeginLoc(), TaskLHSs, TaskRHSs, Data); |
| 1284 | const Expr *TaskRedRef = nullptr; |
| 1285 | switch (D.getDirectiveKind()) { |
| 1286 | case OMPD_parallel: |
| 1287 | TaskRedRef = cast<OMPParallelDirective>(D).getTaskReductionRefExpr(); |
| 1288 | break; |
| 1289 | case OMPD_for: |
| 1290 | TaskRedRef = cast<OMPForDirective>(D).getTaskReductionRefExpr(); |
| 1291 | break; |
| 1292 | case OMPD_sections: |
| 1293 | TaskRedRef = cast<OMPSectionsDirective>(D).getTaskReductionRefExpr(); |
| 1294 | break; |
| 1295 | case OMPD_parallel_for: |
| 1296 | TaskRedRef = cast<OMPParallelForDirective>(D).getTaskReductionRefExpr(); |
| 1297 | break; |
| 1298 | case OMPD_parallel_master: |
| 1299 | TaskRedRef = |
| 1300 | cast<OMPParallelMasterDirective>(D).getTaskReductionRefExpr(); |
| 1301 | break; |
| 1302 | case OMPD_parallel_sections: |
| 1303 | TaskRedRef = |
| 1304 | cast<OMPParallelSectionsDirective>(D).getTaskReductionRefExpr(); |
| 1305 | break; |
| 1306 | case OMPD_target_parallel: |
| 1307 | TaskRedRef = |
| 1308 | cast<OMPTargetParallelDirective>(D).getTaskReductionRefExpr(); |
| 1309 | break; |
| 1310 | case OMPD_target_parallel_for: |
| 1311 | TaskRedRef = |
| 1312 | cast<OMPTargetParallelForDirective>(D).getTaskReductionRefExpr(); |
| 1313 | break; |
| 1314 | case OMPD_distribute_parallel_for: |
| 1315 | TaskRedRef = |
| 1316 | cast<OMPDistributeParallelForDirective>(D).getTaskReductionRefExpr(); |
| 1317 | break; |
| 1318 | case OMPD_teams_distribute_parallel_for: |
| 1319 | TaskRedRef = cast<OMPTeamsDistributeParallelForDirective>(D) |
| 1320 | .getTaskReductionRefExpr(); |
| 1321 | break; |
| 1322 | case OMPD_target_teams_distribute_parallel_for: |
| 1323 | TaskRedRef = cast<OMPTargetTeamsDistributeParallelForDirective>(D) |
| 1324 | .getTaskReductionRefExpr(); |
| 1325 | break; |
| 1326 | case OMPD_simd: |
| 1327 | case OMPD_for_simd: |
| 1328 | case OMPD_section: |
| 1329 | case OMPD_single: |
| 1330 | case OMPD_master: |
| 1331 | case OMPD_critical: |
| 1332 | case OMPD_parallel_for_simd: |
| 1333 | case OMPD_task: |
| 1334 | case OMPD_taskyield: |
| 1335 | case OMPD_barrier: |
| 1336 | case OMPD_taskwait: |
| 1337 | case OMPD_taskgroup: |
| 1338 | case OMPD_flush: |
| 1339 | case OMPD_depobj: |
| 1340 | case OMPD_scan: |
| 1341 | case OMPD_ordered: |
| 1342 | case OMPD_atomic: |
| 1343 | case OMPD_teams: |
| 1344 | case OMPD_target: |
| 1345 | case OMPD_cancellation_point: |
| 1346 | case OMPD_cancel: |
| 1347 | case OMPD_target_data: |
| 1348 | case OMPD_target_enter_data: |
| 1349 | case OMPD_target_exit_data: |
| 1350 | case OMPD_taskloop: |
| 1351 | case OMPD_taskloop_simd: |
| 1352 | case OMPD_master_taskloop: |
| 1353 | case OMPD_master_taskloop_simd: |
| 1354 | case OMPD_parallel_master_taskloop: |
| 1355 | case OMPD_parallel_master_taskloop_simd: |
| 1356 | case OMPD_distribute: |
| 1357 | case OMPD_target_update: |
| 1358 | case OMPD_distribute_parallel_for_simd: |
| 1359 | case OMPD_distribute_simd: |
| 1360 | case OMPD_target_parallel_for_simd: |
| 1361 | case OMPD_target_simd: |
| 1362 | case OMPD_teams_distribute: |
| 1363 | case OMPD_teams_distribute_simd: |
| 1364 | case OMPD_teams_distribute_parallel_for_simd: |
| 1365 | case OMPD_target_teams: |
| 1366 | case OMPD_target_teams_distribute: |
| 1367 | case OMPD_target_teams_distribute_parallel_for_simd: |
| 1368 | case OMPD_target_teams_distribute_simd: |
| 1369 | case OMPD_declare_target: |
| 1370 | case OMPD_end_declare_target: |
| 1371 | case OMPD_threadprivate: |
| 1372 | case OMPD_allocate: |
| 1373 | case OMPD_declare_reduction: |
| 1374 | case OMPD_declare_mapper: |
| 1375 | case OMPD_declare_simd: |
| 1376 | case OMPD_requires: |
| 1377 | case OMPD_declare_variant: |
| 1378 | case OMPD_begin_declare_variant: |
| 1379 | case OMPD_end_declare_variant: |
| 1380 | case OMPD_unknown: |
| 1381 | default: |
| 1382 | llvm_unreachable("Enexpected directive with task reductions." ); |
| 1383 | } |
| 1384 | |
| 1385 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(TaskRedRef)->getDecl()); |
| 1386 | EmitVarDecl(*VD); |
| 1387 | EmitStoreOfScalar(ReductionDesc, GetAddrOfLocalVar(VD), |
| 1388 | /*Volatile=*/false, TaskRedRef->getType()); |
| 1389 | } |
| 1390 | } |
| 1391 | |
| 1392 | void CodeGenFunction::EmitOMPReductionClauseFinal( |
| 1393 | const OMPExecutableDirective &D, const OpenMPDirectiveKind ReductionKind) { |
| 1394 | if (!HaveInsertPoint()) |
| 1395 | return; |
| 1396 | llvm::SmallVector<const Expr *, 8> Privates; |
| 1397 | llvm::SmallVector<const Expr *, 8> LHSExprs; |
| 1398 | llvm::SmallVector<const Expr *, 8> RHSExprs; |
| 1399 | llvm::SmallVector<const Expr *, 8> ReductionOps; |
| 1400 | bool HasAtLeastOneReduction = false; |
| 1401 | bool IsReductionWithTaskMod = false; |
| 1402 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1403 | // Do not emit for inscan reductions. |
| 1404 | if (C->getModifier() == OMPC_REDUCTION_inscan) |
| 1405 | continue; |
| 1406 | HasAtLeastOneReduction = true; |
| 1407 | Privates.append(C->privates().begin(), C->privates().end()); |
| 1408 | LHSExprs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 1409 | RHSExprs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 1410 | ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end()); |
| 1411 | IsReductionWithTaskMod = |
| 1412 | IsReductionWithTaskMod || C->getModifier() == OMPC_REDUCTION_task; |
| 1413 | } |
| 1414 | if (HasAtLeastOneReduction) { |
| 1415 | if (IsReductionWithTaskMod) { |
| 1416 | CGM.getOpenMPRuntime().emitTaskReductionFini( |
| 1417 | *this, D.getBeginLoc(), |
| 1418 | isOpenMPWorksharingDirective(D.getDirectiveKind())); |
| 1419 | } |
| 1420 | bool WithNowait = D.getSingleClause<OMPNowaitClause>() || |
| 1421 | isOpenMPParallelDirective(D.getDirectiveKind()) || |
| 1422 | ReductionKind == OMPD_simd; |
| 1423 | bool SimpleReduction = ReductionKind == OMPD_simd; |
| 1424 | // Emit nowait reduction if nowait clause is present or directive is a |
| 1425 | // parallel directive (it always has implicit barrier). |
| 1426 | CGM.getOpenMPRuntime().emitReduction( |
| 1427 | *this, D.getEndLoc(), Privates, LHSExprs, RHSExprs, ReductionOps, |
| 1428 | {WithNowait, SimpleReduction, ReductionKind}); |
| 1429 | } |
| 1430 | } |
| 1431 | |
| 1432 | static void emitPostUpdateForReductionClause( |
| 1433 | CodeGenFunction &CGF, const OMPExecutableDirective &D, |
| 1434 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 1435 | if (!CGF.HaveInsertPoint()) |
| 1436 | return; |
| 1437 | llvm::BasicBlock *DoneBB = nullptr; |
| 1438 | for (const auto *C : D.getClausesOfKind<OMPReductionClause>()) { |
| 1439 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) { |
| 1440 | if (!DoneBB) { |
| 1441 | if (llvm::Value *Cond = CondGen(CGF)) { |
| 1442 | // If the first post-update expression is found, emit conditional |
| 1443 | // block if it was requested. |
| 1444 | llvm::BasicBlock *ThenBB = CGF.createBasicBlock(".omp.reduction.pu" ); |
| 1445 | DoneBB = CGF.createBasicBlock(".omp.reduction.pu.done" ); |
| 1446 | CGF.Builder.CreateCondBr(Cond, ThenBB, DoneBB); |
| 1447 | CGF.EmitBlock(ThenBB); |
| 1448 | } |
| 1449 | } |
| 1450 | CGF.EmitIgnoredExpr(PostUpdate); |
| 1451 | } |
| 1452 | } |
| 1453 | if (DoneBB) |
| 1454 | CGF.EmitBlock(DoneBB, /*IsFinished=*/true); |
| 1455 | } |
| 1456 | |
| 1457 | namespace { |
| 1458 | /// Codegen lambda for appending distribute lower and upper bounds to outlined |
| 1459 | /// parallel function. This is necessary for combined constructs such as |
| 1460 | /// 'distribute parallel for' |
| 1461 | typedef llvm::function_ref<void(CodeGenFunction &, |
| 1462 | const OMPExecutableDirective &, |
| 1463 | llvm::SmallVectorImpl<llvm::Value *> &)> |
| 1464 | CodeGenBoundParametersTy; |
| 1465 | } // anonymous namespace |
| 1466 | |
| 1467 | static void |
| 1468 | checkForLastprivateConditionalUpdate(CodeGenFunction &CGF, |
| 1469 | const OMPExecutableDirective &S) { |
| 1470 | if (CGF.getLangOpts().OpenMP < 50) |
| 1471 | return; |
| 1472 | llvm::DenseSet<CanonicalDeclPtr<const VarDecl>> PrivateDecls; |
| 1473 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 1474 | for (const Expr *Ref : C->varlists()) { |
| 1475 | if (!Ref->getType()->isScalarType()) |
| 1476 | continue; |
| 1477 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); |
| 1478 | if (!DRE) |
| 1479 | continue; |
| 1480 | PrivateDecls.insert(cast<VarDecl>(DRE->getDecl())); |
| 1481 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, Ref); |
| 1482 | } |
| 1483 | } |
| 1484 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { |
| 1485 | for (const Expr *Ref : C->varlists()) { |
| 1486 | if (!Ref->getType()->isScalarType()) |
| 1487 | continue; |
| 1488 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); |
| 1489 | if (!DRE) |
| 1490 | continue; |
| 1491 | PrivateDecls.insert(cast<VarDecl>(DRE->getDecl())); |
| 1492 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, Ref); |
| 1493 | } |
| 1494 | } |
| 1495 | for (const auto *C : S.getClausesOfKind<OMPLinearClause>()) { |
| 1496 | for (const Expr *Ref : C->varlists()) { |
| 1497 | if (!Ref->getType()->isScalarType()) |
| 1498 | continue; |
| 1499 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); |
| 1500 | if (!DRE) |
| 1501 | continue; |
| 1502 | PrivateDecls.insert(cast<VarDecl>(DRE->getDecl())); |
| 1503 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, Ref); |
| 1504 | } |
| 1505 | } |
| 1506 | // Privates should ne analyzed since they are not captured at all. |
| 1507 | // Task reductions may be skipped - tasks are ignored. |
| 1508 | // Firstprivates do not return value but may be passed by reference - no need |
| 1509 | // to check for updated lastprivate conditional. |
| 1510 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 1511 | for (const Expr *Ref : C->varlists()) { |
| 1512 | if (!Ref->getType()->isScalarType()) |
| 1513 | continue; |
| 1514 | const auto *DRE = dyn_cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); |
| 1515 | if (!DRE) |
| 1516 | continue; |
| 1517 | PrivateDecls.insert(cast<VarDecl>(DRE->getDecl())); |
| 1518 | } |
| 1519 | } |
| 1520 | CGF.CGM.getOpenMPRuntime().checkAndEmitSharedLastprivateConditional( |
| 1521 | CGF, S, PrivateDecls); |
| 1522 | } |
| 1523 | |
| 1524 | static void emitCommonOMPParallelDirective( |
| 1525 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 1526 | OpenMPDirectiveKind InnermostKind, const RegionCodeGenTy &CodeGen, |
| 1527 | const CodeGenBoundParametersTy &CodeGenBoundParameters) { |
| 1528 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel); |
| 1529 | llvm::Function *OutlinedFn = |
| 1530 | CGF.CGM.getOpenMPRuntime().emitParallelOutlinedFunction( |
| 1531 | S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen); |
| 1532 | if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) { |
| 1533 | CodeGenFunction::RunCleanupsScope NumThreadsScope(CGF); |
| 1534 | llvm::Value *NumThreads = |
| 1535 | CGF.EmitScalarExpr(NumThreadsClause->getNumThreads(), |
| 1536 | /*IgnoreResultAssign=*/true); |
| 1537 | CGF.CGM.getOpenMPRuntime().emitNumThreadsClause( |
| 1538 | CGF, NumThreads, NumThreadsClause->getBeginLoc()); |
| 1539 | } |
| 1540 | if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) { |
| 1541 | CodeGenFunction::RunCleanupsScope ProcBindScope(CGF); |
| 1542 | CGF.CGM.getOpenMPRuntime().emitProcBindClause( |
| 1543 | CGF, ProcBindClause->getProcBindKind(), ProcBindClause->getBeginLoc()); |
| 1544 | } |
| 1545 | const Expr *IfCond = nullptr; |
| 1546 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 1547 | if (C->getNameModifier() == OMPD_unknown || |
| 1548 | C->getNameModifier() == OMPD_parallel) { |
| 1549 | IfCond = C->getCondition(); |
| 1550 | break; |
| 1551 | } |
| 1552 | } |
| 1553 | |
| 1554 | OMPParallelScope Scope(CGF, S); |
| 1555 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 1556 | // Combining 'distribute' with 'for' requires sharing each 'distribute' chunk |
| 1557 | // lower and upper bounds with the pragma 'for' chunking mechanism. |
| 1558 | // The following lambda takes care of appending the lower and upper bound |
| 1559 | // parameters when necessary |
| 1560 | CodeGenBoundParameters(CGF, S, CapturedVars); |
| 1561 | CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars); |
| 1562 | CGF.CGM.getOpenMPRuntime().emitParallelCall(CGF, S.getBeginLoc(), OutlinedFn, |
| 1563 | CapturedVars, IfCond); |
| 1564 | } |
| 1565 | |
| 1566 | static bool isAllocatableDecl(const VarDecl *VD) { |
| 1567 | const VarDecl *CVD = VD->getCanonicalDecl(); |
| 1568 | if (!CVD->hasAttr<OMPAllocateDeclAttr>()) |
| 1569 | return false; |
| 1570 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); |
| 1571 | // Use the default allocation. |
| 1572 | return !((AA->getAllocatorType() == OMPAllocateDeclAttr::OMPDefaultMemAlloc || |
| 1573 | AA->getAllocatorType() == OMPAllocateDeclAttr::OMPNullMemAlloc) && |
| 1574 | !AA->getAllocator()); |
| 1575 | } |
| 1576 | |
| 1577 | static void emitEmptyBoundParameters(CodeGenFunction &, |
| 1578 | const OMPExecutableDirective &, |
| 1579 | llvm::SmallVectorImpl<llvm::Value *> &) {} |
| 1580 | |
| 1581 | Address CodeGenFunction::OMPBuilderCBHelpers::getAddressOfLocalVariable( |
| 1582 | CodeGenFunction &CGF, const VarDecl *VD) { |
| 1583 | CodeGenModule &CGM = CGF.CGM; |
| 1584 | auto &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 1585 | |
| 1586 | if (!VD) |
| 1587 | return Address::invalid(); |
| 1588 | const VarDecl *CVD = VD->getCanonicalDecl(); |
| 1589 | if (!isAllocatableDecl(CVD)) |
| 1590 | return Address::invalid(); |
| 1591 | llvm::Value *Size; |
| 1592 | CharUnits Align = CGM.getContext().getDeclAlign(CVD); |
| 1593 | if (CVD->getType()->isVariablyModifiedType()) { |
| 1594 | Size = CGF.getTypeSize(CVD->getType()); |
| 1595 | // Align the size: ((size + align - 1) / align) * align |
| 1596 | Size = CGF.Builder.CreateNUWAdd( |
| 1597 | Size, CGM.getSize(Align - CharUnits::fromQuantity(1))); |
| 1598 | Size = CGF.Builder.CreateUDiv(Size, CGM.getSize(Align)); |
| 1599 | Size = CGF.Builder.CreateNUWMul(Size, CGM.getSize(Align)); |
| 1600 | } else { |
| 1601 | CharUnits Sz = CGM.getContext().getTypeSizeInChars(CVD->getType()); |
| 1602 | Size = CGM.getSize(Sz.alignTo(Align)); |
| 1603 | } |
| 1604 | |
| 1605 | const auto *AA = CVD->getAttr<OMPAllocateDeclAttr>(); |
| 1606 | assert(AA->getAllocator() && |
| 1607 | "Expected allocator expression for non-default allocator." ); |
| 1608 | llvm::Value *Allocator = CGF.EmitScalarExpr(AA->getAllocator()); |
| 1609 | // According to the standard, the original allocator type is a enum (integer). |
| 1610 | // Convert to pointer type, if required. |
| 1611 | if (Allocator->getType()->isIntegerTy()) |
| 1612 | Allocator = CGF.Builder.CreateIntToPtr(Allocator, CGM.VoidPtrTy); |
| 1613 | else if (Allocator->getType()->isPointerTy()) |
| 1614 | Allocator = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Allocator, |
| 1615 | CGM.VoidPtrTy); |
| 1616 | |
| 1617 | llvm::Value *Addr = OMPBuilder.createOMPAlloc( |
| 1618 | CGF.Builder, Size, Allocator, |
| 1619 | getNameWithSeparators({CVD->getName(), ".void.addr" }, "." , "." )); |
| 1620 | llvm::CallInst *FreeCI = |
| 1621 | OMPBuilder.createOMPFree(CGF.Builder, Addr, Allocator); |
| 1622 | |
| 1623 | CGF.EHStack.pushCleanup<OMPAllocateCleanupTy>(NormalAndEHCleanup, FreeCI); |
| 1624 | Addr = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast( |
| 1625 | Addr, |
| 1626 | CGF.ConvertTypeForMem(CGM.getContext().getPointerType(CVD->getType())), |
| 1627 | getNameWithSeparators({CVD->getName(), ".addr" }, "." , "." )); |
| 1628 | return Address(Addr, Align); |
| 1629 | } |
| 1630 | |
| 1631 | Address CodeGenFunction::OMPBuilderCBHelpers::getAddrOfThreadPrivate( |
| 1632 | CodeGenFunction &CGF, const VarDecl *VD, Address VDAddr, |
| 1633 | SourceLocation Loc) { |
| 1634 | CodeGenModule &CGM = CGF.CGM; |
| 1635 | if (CGM.getLangOpts().OpenMPUseTLS && |
| 1636 | CGM.getContext().getTargetInfo().isTLSSupported()) |
| 1637 | return VDAddr; |
| 1638 | |
| 1639 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 1640 | |
| 1641 | llvm::Type *VarTy = VDAddr.getElementType(); |
| 1642 | llvm::Value *Data = |
| 1643 | CGF.Builder.CreatePointerCast(VDAddr.getPointer(), CGM.Int8PtrTy); |
| 1644 | llvm::ConstantInt *Size = CGM.getSize(CGM.GetTargetTypeStoreSize(VarTy)); |
| 1645 | std::string Suffix = getNameWithSeparators({"cache" , "" }); |
| 1646 | llvm::Twine CacheName = Twine(CGM.getMangledName(VD)).concat(Suffix); |
| 1647 | |
| 1648 | llvm::CallInst *ThreadPrivateCacheCall = |
| 1649 | OMPBuilder.createCachedThreadPrivate(CGF.Builder, Data, Size, CacheName); |
| 1650 | |
| 1651 | return Address(ThreadPrivateCacheCall, VDAddr.getAlignment()); |
| 1652 | } |
| 1653 | |
| 1654 | std::string CodeGenFunction::OMPBuilderCBHelpers::getNameWithSeparators( |
| 1655 | ArrayRef<StringRef> Parts, StringRef FirstSeparator, StringRef Separator) { |
| 1656 | SmallString<128> Buffer; |
| 1657 | llvm::raw_svector_ostream OS(Buffer); |
| 1658 | StringRef Sep = FirstSeparator; |
| 1659 | for (StringRef Part : Parts) { |
| 1660 | OS << Sep << Part; |
| 1661 | Sep = Separator; |
| 1662 | } |
| 1663 | return OS.str().str(); |
| 1664 | } |
| 1665 | void CodeGenFunction::EmitOMPParallelDirective(const OMPParallelDirective &S) { |
| 1666 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 1667 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 1668 | // Check if we have any if clause associated with the directive. |
| 1669 | llvm::Value *IfCond = nullptr; |
| 1670 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 1671 | IfCond = EmitScalarExpr(C->getCondition(), |
| 1672 | /*IgnoreResultAssign=*/true); |
| 1673 | |
| 1674 | llvm::Value *NumThreads = nullptr; |
| 1675 | if (const auto *NumThreadsClause = S.getSingleClause<OMPNumThreadsClause>()) |
| 1676 | NumThreads = EmitScalarExpr(NumThreadsClause->getNumThreads(), |
| 1677 | /*IgnoreResultAssign=*/true); |
| 1678 | |
| 1679 | ProcBindKind ProcBind = OMP_PROC_BIND_default; |
| 1680 | if (const auto *ProcBindClause = S.getSingleClause<OMPProcBindClause>()) |
| 1681 | ProcBind = ProcBindClause->getProcBindKind(); |
| 1682 | |
| 1683 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 1684 | |
| 1685 | // The cleanup callback that finalizes all variabels at the given location, |
| 1686 | // thus calls destructors etc. |
| 1687 | auto FiniCB = [this](InsertPointTy IP) { |
| 1688 | OMPBuilderCBHelpers::FinalizeOMPRegion(*this, IP); |
| 1689 | }; |
| 1690 | |
| 1691 | // Privatization callback that performs appropriate action for |
| 1692 | // shared/private/firstprivate/lastprivate/copyin/... variables. |
| 1693 | // |
| 1694 | // TODO: This defaults to shared right now. |
| 1695 | auto PrivCB = [](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, |
| 1696 | llvm::Value &, llvm::Value &Val, llvm::Value *&ReplVal) { |
| 1697 | // The next line is appropriate only for variables (Val) with the |
| 1698 | // data-sharing attribute "shared". |
| 1699 | ReplVal = &Val; |
| 1700 | |
| 1701 | return CodeGenIP; |
| 1702 | }; |
| 1703 | |
| 1704 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel); |
| 1705 | const Stmt *ParallelRegionBodyStmt = CS->getCapturedStmt(); |
| 1706 | |
| 1707 | auto BodyGenCB = [ParallelRegionBodyStmt, |
| 1708 | this](InsertPointTy AllocaIP, InsertPointTy CodeGenIP, |
| 1709 | llvm::BasicBlock &ContinuationBB) { |
| 1710 | OMPBuilderCBHelpers::OutlinedRegionBodyRAII ORB(*this, AllocaIP, |
| 1711 | ContinuationBB); |
| 1712 | OMPBuilderCBHelpers::EmitOMPRegionBody(*this, ParallelRegionBodyStmt, |
| 1713 | CodeGenIP, ContinuationBB); |
| 1714 | }; |
| 1715 | |
| 1716 | CGCapturedStmtInfo CGSI(*CS, CR_OpenMP); |
| 1717 | CodeGenFunction::CGCapturedStmtRAII CapInfoRAII(*this, &CGSI); |
| 1718 | llvm::OpenMPIRBuilder::InsertPointTy AllocaIP( |
| 1719 | AllocaInsertPt->getParent(), AllocaInsertPt->getIterator()); |
| 1720 | Builder.restoreIP( |
| 1721 | OMPBuilder.createParallel(Builder, AllocaIP, BodyGenCB, PrivCB, FiniCB, |
| 1722 | IfCond, NumThreads, ProcBind, S.hasCancel())); |
| 1723 | return; |
| 1724 | } |
| 1725 | |
| 1726 | // Emit parallel region as a standalone region. |
| 1727 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 1728 | Action.Enter(CGF); |
| 1729 | OMPPrivateScope PrivateScope(CGF); |
| 1730 | bool Copyins = CGF.EmitOMPCopyinClause(S); |
| 1731 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 1732 | if (Copyins) { |
| 1733 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 1734 | // propagation master's thread values of threadprivate variables to local |
| 1735 | // instances of that variables of all other implicit threads. |
| 1736 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 1737 | CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false, |
| 1738 | /*ForceSimpleCall=*/true); |
| 1739 | } |
| 1740 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 1741 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 1742 | (void)PrivateScope.Privatize(); |
| 1743 | CGF.EmitStmt(S.getCapturedStmt(OMPD_parallel)->getCapturedStmt()); |
| 1744 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel); |
| 1745 | }; |
| 1746 | { |
| 1747 | auto LPCRegion = |
| 1748 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 1749 | emitCommonOMPParallelDirective(*this, S, OMPD_parallel, CodeGen, |
| 1750 | emitEmptyBoundParameters); |
| 1751 | emitPostUpdateForReductionClause(*this, S, |
| 1752 | [](CodeGenFunction &) { return nullptr; }); |
| 1753 | } |
| 1754 | // Check for outer lastprivate conditional update. |
| 1755 | checkForLastprivateConditionalUpdate(*this, S); |
| 1756 | } |
| 1757 | |
| 1758 | static void emitBody(CodeGenFunction &CGF, const Stmt *S, const Stmt *NextLoop, |
| 1759 | int MaxLevel, int Level = 0) { |
| 1760 | assert(Level < MaxLevel && "Too deep lookup during loop body codegen." ); |
| 1761 | const Stmt *SimplifiedS = S->IgnoreContainers(); |
| 1762 | if (const auto *CS = dyn_cast<CompoundStmt>(SimplifiedS)) { |
| 1763 | PrettyStackTraceLoc CrashInfo( |
| 1764 | CGF.getContext().getSourceManager(), CS->getLBracLoc(), |
| 1765 | "LLVM IR generation of compound statement ('{}')" ); |
| 1766 | |
| 1767 | // Keep track of the current cleanup stack depth, including debug scopes. |
| 1768 | CodeGenFunction::LexicalScope Scope(CGF, S->getSourceRange()); |
| 1769 | for (const Stmt *CurStmt : CS->body()) |
| 1770 | emitBody(CGF, CurStmt, NextLoop, MaxLevel, Level); |
| 1771 | return; |
| 1772 | } |
| 1773 | if (SimplifiedS == NextLoop) { |
| 1774 | if (const auto *For = dyn_cast<ForStmt>(SimplifiedS)) { |
| 1775 | S = For->getBody(); |
| 1776 | } else { |
| 1777 | assert(isa<CXXForRangeStmt>(SimplifiedS) && |
| 1778 | "Expected canonical for loop or range-based for loop." ); |
| 1779 | const auto *CXXFor = cast<CXXForRangeStmt>(SimplifiedS); |
| 1780 | CGF.EmitStmt(CXXFor->getLoopVarStmt()); |
| 1781 | S = CXXFor->getBody(); |
| 1782 | } |
| 1783 | if (Level + 1 < MaxLevel) { |
| 1784 | NextLoop = OMPLoopDirective::tryToFindNextInnerLoop( |
| 1785 | S, /*TryImperfectlyNestedLoops=*/true); |
| 1786 | emitBody(CGF, S, NextLoop, MaxLevel, Level + 1); |
| 1787 | return; |
| 1788 | } |
| 1789 | } |
| 1790 | CGF.EmitStmt(S); |
| 1791 | } |
| 1792 | |
| 1793 | void CodeGenFunction::EmitOMPLoopBody(const OMPLoopDirective &D, |
| 1794 | JumpDest LoopExit) { |
| 1795 | RunCleanupsScope BodyScope(*this); |
| 1796 | // Update counters values on current iteration. |
| 1797 | for (const Expr *UE : D.updates()) |
| 1798 | EmitIgnoredExpr(UE); |
| 1799 | // Update the linear variables. |
| 1800 | // In distribute directives only loop counters may be marked as linear, no |
| 1801 | // need to generate the code for them. |
| 1802 | if (!isOpenMPDistributeDirective(D.getDirectiveKind())) { |
| 1803 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 1804 | for (const Expr *UE : C->updates()) |
| 1805 | EmitIgnoredExpr(UE); |
| 1806 | } |
| 1807 | } |
| 1808 | |
| 1809 | // On a continue in the body, jump to the end. |
| 1810 | JumpDest Continue = getJumpDestInCurrentScope("omp.body.continue" ); |
| 1811 | BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); |
| 1812 | for (const Expr *E : D.finals_conditions()) { |
| 1813 | if (!E) |
| 1814 | continue; |
| 1815 | // Check that loop counter in non-rectangular nest fits into the iteration |
| 1816 | // space. |
| 1817 | llvm::BasicBlock *NextBB = createBasicBlock("omp.body.next" ); |
| 1818 | EmitBranchOnBoolExpr(E, NextBB, Continue.getBlock(), |
| 1819 | getProfileCount(D.getBody())); |
| 1820 | EmitBlock(NextBB); |
| 1821 | } |
| 1822 | |
| 1823 | OMPPrivateScope InscanScope(*this); |
| 1824 | EmitOMPReductionClauseInit(D, InscanScope, /*ForInscan=*/true); |
| 1825 | bool IsInscanRegion = InscanScope.Privatize(); |
| 1826 | if (IsInscanRegion) { |
| 1827 | // Need to remember the block before and after scan directive |
| 1828 | // to dispatch them correctly depending on the clause used in |
| 1829 | // this directive, inclusive or exclusive. For inclusive scan the natural |
| 1830 | // order of the blocks is used, for exclusive clause the blocks must be |
| 1831 | // executed in reverse order. |
| 1832 | OMPBeforeScanBlock = createBasicBlock("omp.before.scan.bb" ); |
| 1833 | OMPAfterScanBlock = createBasicBlock("omp.after.scan.bb" ); |
| 1834 | // No need to allocate inscan exit block, in simd mode it is selected in the |
| 1835 | // codegen for the scan directive. |
| 1836 | if (D.getDirectiveKind() != OMPD_simd && !getLangOpts().OpenMPSimd) |
| 1837 | OMPScanExitBlock = createBasicBlock("omp.exit.inscan.bb" ); |
| 1838 | OMPScanDispatch = createBasicBlock("omp.inscan.dispatch" ); |
| 1839 | EmitBranch(OMPScanDispatch); |
| 1840 | EmitBlock(OMPBeforeScanBlock); |
| 1841 | } |
| 1842 | |
| 1843 | // Emit loop variables for C++ range loops. |
| 1844 | const Stmt *Body = |
| 1845 | D.getInnermostCapturedStmt()->getCapturedStmt()->IgnoreContainers(); |
| 1846 | // Emit loop body. |
| 1847 | emitBody(*this, Body, |
| 1848 | OMPLoopDirective::tryToFindNextInnerLoop( |
| 1849 | Body, /*TryImperfectlyNestedLoops=*/true), |
| 1850 | D.getCollapsedNumber()); |
| 1851 | |
| 1852 | // Jump to the dispatcher at the end of the loop body. |
| 1853 | if (IsInscanRegion) |
| 1854 | EmitBranch(OMPScanExitBlock); |
| 1855 | |
| 1856 | // The end (updates/cleanups). |
| 1857 | EmitBlock(Continue.getBlock()); |
| 1858 | BreakContinueStack.pop_back(); |
| 1859 | } |
| 1860 | |
| 1861 | void CodeGenFunction::EmitOMPInnerLoop( |
| 1862 | const OMPExecutableDirective &S, bool RequiresCleanup, const Expr *LoopCond, |
| 1863 | const Expr *IncExpr, |
| 1864 | const llvm::function_ref<void(CodeGenFunction &)> BodyGen, |
| 1865 | const llvm::function_ref<void(CodeGenFunction &)> PostIncGen) { |
| 1866 | auto LoopExit = getJumpDestInCurrentScope("omp.inner.for.end" ); |
| 1867 | |
| 1868 | // Start the loop with a block that tests the condition. |
| 1869 | auto CondBlock = createBasicBlock("omp.inner.for.cond" ); |
| 1870 | EmitBlock(CondBlock); |
| 1871 | const SourceRange R = S.getSourceRange(); |
| 1872 | |
| 1873 | // If attributes are attached, push to the basic block with them. |
| 1874 | const auto &OMPED = cast<OMPExecutableDirective>(S); |
| 1875 | const CapturedStmt *ICS = OMPED.getInnermostCapturedStmt(); |
| 1876 | const Stmt *SS = ICS->getCapturedStmt(); |
| 1877 | const AttributedStmt *AS = dyn_cast_or_null<AttributedStmt>(SS); |
| 1878 | if (AS) |
| 1879 | LoopStack.push(CondBlock, CGM.getContext(), CGM.getCodeGenOpts(), |
| 1880 | AS->getAttrs(), SourceLocToDebugLoc(R.getBegin()), |
| 1881 | SourceLocToDebugLoc(R.getEnd())); |
| 1882 | else |
| 1883 | LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()), |
| 1884 | SourceLocToDebugLoc(R.getEnd())); |
| 1885 | |
| 1886 | // If there are any cleanups between here and the loop-exit scope, |
| 1887 | // create a block to stage a loop exit along. |
| 1888 | llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 1889 | if (RequiresCleanup) |
| 1890 | ExitBlock = createBasicBlock("omp.inner.for.cond.cleanup" ); |
| 1891 | |
| 1892 | llvm::BasicBlock *LoopBody = createBasicBlock("omp.inner.for.body" ); |
| 1893 | |
| 1894 | // Emit condition. |
| 1895 | EmitBranchOnBoolExpr(LoopCond, LoopBody, ExitBlock, getProfileCount(&S)); |
| 1896 | if (ExitBlock != LoopExit.getBlock()) { |
| 1897 | EmitBlock(ExitBlock); |
| 1898 | EmitBranchThroughCleanup(LoopExit); |
| 1899 | } |
| 1900 | |
| 1901 | EmitBlock(LoopBody); |
| 1902 | incrementProfileCounter(&S); |
| 1903 | |
| 1904 | // Create a block for the increment. |
| 1905 | JumpDest Continue = getJumpDestInCurrentScope("omp.inner.for.inc" ); |
| 1906 | BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); |
| 1907 | |
| 1908 | BodyGen(*this); |
| 1909 | |
| 1910 | // Emit "IV = IV + 1" and a back-edge to the condition block. |
| 1911 | EmitBlock(Continue.getBlock()); |
| 1912 | EmitIgnoredExpr(IncExpr); |
| 1913 | PostIncGen(*this); |
| 1914 | BreakContinueStack.pop_back(); |
| 1915 | EmitBranch(CondBlock); |
| 1916 | LoopStack.pop(); |
| 1917 | // Emit the fall-through block. |
| 1918 | EmitBlock(LoopExit.getBlock()); |
| 1919 | } |
| 1920 | |
| 1921 | bool CodeGenFunction::EmitOMPLinearClauseInit(const OMPLoopDirective &D) { |
| 1922 | if (!HaveInsertPoint()) |
| 1923 | return false; |
| 1924 | // Emit inits for the linear variables. |
| 1925 | bool HasLinears = false; |
| 1926 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 1927 | for (const Expr *Init : C->inits()) { |
| 1928 | HasLinears = true; |
| 1929 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(Init)->getDecl()); |
| 1930 | if (const auto *Ref = |
| 1931 | dyn_cast<DeclRefExpr>(VD->getInit()->IgnoreImpCasts())) { |
| 1932 | AutoVarEmission Emission = EmitAutoVarAlloca(*VD); |
| 1933 | const auto *OrigVD = cast<VarDecl>(Ref->getDecl()); |
| 1934 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 1935 | CapturedStmtInfo->lookup(OrigVD) != nullptr, |
| 1936 | VD->getInit()->getType(), VK_LValue, |
| 1937 | VD->getInit()->getExprLoc()); |
| 1938 | EmitExprAsInit(&DRE, VD, MakeAddrLValue(Emission.getAllocatedAddress(), |
| 1939 | VD->getType()), |
| 1940 | /*capturedByInit=*/false); |
| 1941 | EmitAutoVarCleanups(Emission); |
| 1942 | } else { |
| 1943 | EmitVarDecl(*VD); |
| 1944 | } |
| 1945 | } |
| 1946 | // Emit the linear steps for the linear clauses. |
| 1947 | // If a step is not constant, it is pre-calculated before the loop. |
| 1948 | if (const auto *CS = cast_or_null<BinaryOperator>(C->getCalcStep())) |
| 1949 | if (const auto *SaveRef = cast<DeclRefExpr>(CS->getLHS())) { |
| 1950 | EmitVarDecl(*cast<VarDecl>(SaveRef->getDecl())); |
| 1951 | // Emit calculation of the linear step. |
| 1952 | EmitIgnoredExpr(CS); |
| 1953 | } |
| 1954 | } |
| 1955 | return HasLinears; |
| 1956 | } |
| 1957 | |
| 1958 | void CodeGenFunction::EmitOMPLinearClauseFinal( |
| 1959 | const OMPLoopDirective &D, |
| 1960 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 1961 | if (!HaveInsertPoint()) |
| 1962 | return; |
| 1963 | llvm::BasicBlock *DoneBB = nullptr; |
| 1964 | // Emit the final values of the linear variables. |
| 1965 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 1966 | auto IC = C->varlist_begin(); |
| 1967 | for (const Expr *F : C->finals()) { |
| 1968 | if (!DoneBB) { |
| 1969 | if (llvm::Value *Cond = CondGen(*this)) { |
| 1970 | // If the first post-update expression is found, emit conditional |
| 1971 | // block if it was requested. |
| 1972 | llvm::BasicBlock *ThenBB = createBasicBlock(".omp.linear.pu" ); |
| 1973 | DoneBB = createBasicBlock(".omp.linear.pu.done" ); |
| 1974 | Builder.CreateCondBr(Cond, ThenBB, DoneBB); |
| 1975 | EmitBlock(ThenBB); |
| 1976 | } |
| 1977 | } |
| 1978 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IC)->getDecl()); |
| 1979 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(OrigVD), |
| 1980 | CapturedStmtInfo->lookup(OrigVD) != nullptr, |
| 1981 | (*IC)->getType(), VK_LValue, (*IC)->getExprLoc()); |
| 1982 | Address OrigAddr = EmitLValue(&DRE).getAddress(*this); |
| 1983 | CodeGenFunction::OMPPrivateScope VarScope(*this); |
| 1984 | VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; }); |
| 1985 | (void)VarScope.Privatize(); |
| 1986 | EmitIgnoredExpr(F); |
| 1987 | ++IC; |
| 1988 | } |
| 1989 | if (const Expr *PostUpdate = C->getPostUpdateExpr()) |
| 1990 | EmitIgnoredExpr(PostUpdate); |
| 1991 | } |
| 1992 | if (DoneBB) |
| 1993 | EmitBlock(DoneBB, /*IsFinished=*/true); |
| 1994 | } |
| 1995 | |
| 1996 | static void emitAlignedClause(CodeGenFunction &CGF, |
| 1997 | const OMPExecutableDirective &D) { |
| 1998 | if (!CGF.HaveInsertPoint()) |
| 1999 | return; |
| 2000 | for (const auto *Clause : D.getClausesOfKind<OMPAlignedClause>()) { |
| 2001 | llvm::APInt ClauseAlignment(64, 0); |
| 2002 | if (const Expr *AlignmentExpr = Clause->getAlignment()) { |
| 2003 | auto *AlignmentCI = |
| 2004 | cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AlignmentExpr)); |
| 2005 | ClauseAlignment = AlignmentCI->getValue(); |
| 2006 | } |
| 2007 | for (const Expr *E : Clause->varlists()) { |
| 2008 | llvm::APInt Alignment(ClauseAlignment); |
| 2009 | if (Alignment == 0) { |
| 2010 | // OpenMP [2.8.1, Description] |
| 2011 | // If no optional parameter is specified, implementation-defined default |
| 2012 | // alignments for SIMD instructions on the target platforms are assumed. |
| 2013 | Alignment = |
| 2014 | CGF.getContext() |
| 2015 | .toCharUnitsFromBits(CGF.getContext().getOpenMPDefaultSimdAlign( |
| 2016 | E->getType()->getPointeeType())) |
| 2017 | .getQuantity(); |
| 2018 | } |
| 2019 | assert((Alignment == 0 || Alignment.isPowerOf2()) && |
| 2020 | "alignment is not power of 2" ); |
| 2021 | if (Alignment != 0) { |
| 2022 | llvm::Value *PtrValue = CGF.EmitScalarExpr(E); |
| 2023 | CGF.emitAlignmentAssumption( |
| 2024 | PtrValue, E, /*No second loc needed*/ SourceLocation(), |
| 2025 | llvm::ConstantInt::get(CGF.getLLVMContext(), Alignment)); |
| 2026 | } |
| 2027 | } |
| 2028 | } |
| 2029 | } |
| 2030 | |
| 2031 | void CodeGenFunction::EmitOMPPrivateLoopCounters( |
| 2032 | const OMPLoopDirective &S, CodeGenFunction::OMPPrivateScope &LoopScope) { |
| 2033 | if (!HaveInsertPoint()) |
| 2034 | return; |
| 2035 | auto I = S.private_counters().begin(); |
| 2036 | for (const Expr *E : S.counters()) { |
| 2037 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 2038 | const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(*I)->getDecl()); |
| 2039 | // Emit var without initialization. |
| 2040 | AutoVarEmission VarEmission = EmitAutoVarAlloca(*PrivateVD); |
| 2041 | EmitAutoVarCleanups(VarEmission); |
| 2042 | LocalDeclMap.erase(PrivateVD); |
| 2043 | (void)LoopScope.addPrivate(VD, [&VarEmission]() { |
| 2044 | return VarEmission.getAllocatedAddress(); |
| 2045 | }); |
| 2046 | if (LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD) || |
| 2047 | VD->hasGlobalStorage()) { |
| 2048 | (void)LoopScope.addPrivate(PrivateVD, [this, VD, E]() { |
| 2049 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(VD), |
| 2050 | LocalDeclMap.count(VD) || CapturedStmtInfo->lookup(VD), |
| 2051 | E->getType(), VK_LValue, E->getExprLoc()); |
| 2052 | return EmitLValue(&DRE).getAddress(*this); |
| 2053 | }); |
| 2054 | } else { |
| 2055 | (void)LoopScope.addPrivate(PrivateVD, [&VarEmission]() { |
| 2056 | return VarEmission.getAllocatedAddress(); |
| 2057 | }); |
| 2058 | } |
| 2059 | ++I; |
| 2060 | } |
| 2061 | // Privatize extra loop counters used in loops for ordered(n) clauses. |
| 2062 | for (const auto *C : S.getClausesOfKind<OMPOrderedClause>()) { |
| 2063 | if (!C->getNumForLoops()) |
| 2064 | continue; |
| 2065 | for (unsigned I = S.getCollapsedNumber(), |
| 2066 | E = C->getLoopNumIterations().size(); |
| 2067 | I < E; ++I) { |
| 2068 | const auto *DRE = cast<DeclRefExpr>(C->getLoopCounter(I)); |
| 2069 | const auto *VD = cast<VarDecl>(DRE->getDecl()); |
| 2070 | // Override only those variables that can be captured to avoid re-emission |
| 2071 | // of the variables declared within the loops. |
| 2072 | if (DRE->refersToEnclosingVariableOrCapture()) { |
| 2073 | (void)LoopScope.addPrivate(VD, [this, DRE, VD]() { |
| 2074 | return CreateMemTemp(DRE->getType(), VD->getName()); |
| 2075 | }); |
| 2076 | } |
| 2077 | } |
| 2078 | } |
| 2079 | } |
| 2080 | |
| 2081 | static void emitPreCond(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2082 | const Expr *Cond, llvm::BasicBlock *TrueBlock, |
| 2083 | llvm::BasicBlock *FalseBlock, uint64_t TrueCount) { |
| 2084 | if (!CGF.HaveInsertPoint()) |
| 2085 | return; |
| 2086 | { |
| 2087 | CodeGenFunction::OMPPrivateScope PreCondScope(CGF); |
| 2088 | CGF.EmitOMPPrivateLoopCounters(S, PreCondScope); |
| 2089 | (void)PreCondScope.Privatize(); |
| 2090 | // Get initial values of real counters. |
| 2091 | for (const Expr *I : S.inits()) { |
| 2092 | CGF.EmitIgnoredExpr(I); |
| 2093 | } |
| 2094 | } |
| 2095 | // Create temp loop control variables with their init values to support |
| 2096 | // non-rectangular loops. |
| 2097 | CodeGenFunction::OMPMapVars PreCondVars; |
| 2098 | for (const Expr * E: S.dependent_counters()) { |
| 2099 | if (!E) |
| 2100 | continue; |
| 2101 | assert(!E->getType().getNonReferenceType()->isRecordType() && |
| 2102 | "dependent counter must not be an iterator." ); |
| 2103 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 2104 | Address CounterAddr = |
| 2105 | CGF.CreateMemTemp(VD->getType().getNonReferenceType()); |
| 2106 | (void)PreCondVars.setVarAddr(CGF, VD, CounterAddr); |
| 2107 | } |
| 2108 | (void)PreCondVars.apply(CGF); |
| 2109 | for (const Expr *E : S.dependent_inits()) { |
| 2110 | if (!E) |
| 2111 | continue; |
| 2112 | CGF.EmitIgnoredExpr(E); |
| 2113 | } |
| 2114 | // Check that loop is executed at least one time. |
| 2115 | CGF.EmitBranchOnBoolExpr(Cond, TrueBlock, FalseBlock, TrueCount); |
| 2116 | PreCondVars.restore(CGF); |
| 2117 | } |
| 2118 | |
| 2119 | void CodeGenFunction::EmitOMPLinearClause( |
| 2120 | const OMPLoopDirective &D, CodeGenFunction::OMPPrivateScope &PrivateScope) { |
| 2121 | if (!HaveInsertPoint()) |
| 2122 | return; |
| 2123 | llvm::DenseSet<const VarDecl *> SIMDLCVs; |
| 2124 | if (isOpenMPSimdDirective(D.getDirectiveKind())) { |
| 2125 | const auto *LoopDirective = cast<OMPLoopDirective>(&D); |
| 2126 | for (const Expr *C : LoopDirective->counters()) { |
| 2127 | SIMDLCVs.insert( |
| 2128 | cast<VarDecl>(cast<DeclRefExpr>(C)->getDecl())->getCanonicalDecl()); |
| 2129 | } |
| 2130 | } |
| 2131 | for (const auto *C : D.getClausesOfKind<OMPLinearClause>()) { |
| 2132 | auto CurPrivate = C->privates().begin(); |
| 2133 | for (const Expr *E : C->varlists()) { |
| 2134 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 2135 | const auto *PrivateVD = |
| 2136 | cast<VarDecl>(cast<DeclRefExpr>(*CurPrivate)->getDecl()); |
| 2137 | if (!SIMDLCVs.count(VD->getCanonicalDecl())) { |
| 2138 | bool IsRegistered = PrivateScope.addPrivate(VD, [this, PrivateVD]() { |
| 2139 | // Emit private VarDecl with copy init. |
| 2140 | EmitVarDecl(*PrivateVD); |
| 2141 | return GetAddrOfLocalVar(PrivateVD); |
| 2142 | }); |
| 2143 | assert(IsRegistered && "linear var already registered as private" ); |
| 2144 | // Silence the warning about unused variable. |
| 2145 | (void)IsRegistered; |
| 2146 | } else { |
| 2147 | EmitVarDecl(*PrivateVD); |
| 2148 | } |
| 2149 | ++CurPrivate; |
| 2150 | } |
| 2151 | } |
| 2152 | } |
| 2153 | |
| 2154 | static void emitSimdlenSafelenClause(CodeGenFunction &CGF, |
| 2155 | const OMPExecutableDirective &D, |
| 2156 | bool IsMonotonic) { |
| 2157 | if (!CGF.HaveInsertPoint()) |
| 2158 | return; |
| 2159 | if (const auto *C = D.getSingleClause<OMPSimdlenClause>()) { |
| 2160 | RValue Len = CGF.EmitAnyExpr(C->getSimdlen(), AggValueSlot::ignored(), |
| 2161 | /*ignoreResult=*/true); |
| 2162 | auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal()); |
| 2163 | CGF.LoopStack.setVectorizeWidth(Val->getZExtValue()); |
| 2164 | // In presence of finite 'safelen', it may be unsafe to mark all |
| 2165 | // the memory instructions parallel, because loop-carried |
| 2166 | // dependences of 'safelen' iterations are possible. |
| 2167 | if (!IsMonotonic) |
| 2168 | CGF.LoopStack.setParallel(!D.getSingleClause<OMPSafelenClause>()); |
| 2169 | } else if (const auto *C = D.getSingleClause<OMPSafelenClause>()) { |
| 2170 | RValue Len = CGF.EmitAnyExpr(C->getSafelen(), AggValueSlot::ignored(), |
| 2171 | /*ignoreResult=*/true); |
| 2172 | auto *Val = cast<llvm::ConstantInt>(Len.getScalarVal()); |
| 2173 | CGF.LoopStack.setVectorizeWidth(Val->getZExtValue()); |
| 2174 | // In presence of finite 'safelen', it may be unsafe to mark all |
| 2175 | // the memory instructions parallel, because loop-carried |
| 2176 | // dependences of 'safelen' iterations are possible. |
| 2177 | CGF.LoopStack.setParallel(/*Enable=*/false); |
| 2178 | } |
| 2179 | } |
| 2180 | |
| 2181 | void CodeGenFunction::EmitOMPSimdInit(const OMPLoopDirective &D, |
| 2182 | bool IsMonotonic) { |
| 2183 | // Walk clauses and process safelen/lastprivate. |
| 2184 | LoopStack.setParallel(!IsMonotonic); |
| 2185 | LoopStack.setVectorizeEnable(); |
| 2186 | emitSimdlenSafelenClause(*this, D, IsMonotonic); |
| 2187 | if (const auto *C = D.getSingleClause<OMPOrderClause>()) |
| 2188 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 2189 | LoopStack.setParallel(/*Enable=*/true); |
| 2190 | if ((D.getDirectiveKind() == OMPD_simd || |
| 2191 | (getLangOpts().OpenMPSimd && |
| 2192 | isOpenMPSimdDirective(D.getDirectiveKind()))) && |
| 2193 | llvm::any_of(D.getClausesOfKind<OMPReductionClause>(), |
| 2194 | [](const OMPReductionClause *C) { |
| 2195 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 2196 | })) |
| 2197 | // Disable parallel access in case of prefix sum. |
| 2198 | LoopStack.setParallel(/*Enable=*/false); |
| 2199 | } |
| 2200 | |
| 2201 | void CodeGenFunction::EmitOMPSimdFinal( |
| 2202 | const OMPLoopDirective &D, |
| 2203 | const llvm::function_ref<llvm::Value *(CodeGenFunction &)> CondGen) { |
| 2204 | if (!HaveInsertPoint()) |
| 2205 | return; |
| 2206 | llvm::BasicBlock *DoneBB = nullptr; |
| 2207 | auto IC = D.counters().begin(); |
| 2208 | auto IPC = D.private_counters().begin(); |
| 2209 | for (const Expr *F : D.finals()) { |
| 2210 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>((*IC))->getDecl()); |
| 2211 | const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>((*IPC))->getDecl()); |
| 2212 | const auto *CED = dyn_cast<OMPCapturedExprDecl>(OrigVD); |
| 2213 | if (LocalDeclMap.count(OrigVD) || CapturedStmtInfo->lookup(OrigVD) || |
| 2214 | OrigVD->hasGlobalStorage() || CED) { |
| 2215 | if (!DoneBB) { |
| 2216 | if (llvm::Value *Cond = CondGen(*this)) { |
| 2217 | // If the first post-update expression is found, emit conditional |
| 2218 | // block if it was requested. |
| 2219 | llvm::BasicBlock *ThenBB = createBasicBlock(".omp.final.then" ); |
| 2220 | DoneBB = createBasicBlock(".omp.final.done" ); |
| 2221 | Builder.CreateCondBr(Cond, ThenBB, DoneBB); |
| 2222 | EmitBlock(ThenBB); |
| 2223 | } |
| 2224 | } |
| 2225 | Address OrigAddr = Address::invalid(); |
| 2226 | if (CED) { |
| 2227 | OrigAddr = |
| 2228 | EmitLValue(CED->getInit()->IgnoreImpCasts()).getAddress(*this); |
| 2229 | } else { |
| 2230 | DeclRefExpr DRE(getContext(), const_cast<VarDecl *>(PrivateVD), |
| 2231 | /*RefersToEnclosingVariableOrCapture=*/false, |
| 2232 | (*IPC)->getType(), VK_LValue, (*IPC)->getExprLoc()); |
| 2233 | OrigAddr = EmitLValue(&DRE).getAddress(*this); |
| 2234 | } |
| 2235 | OMPPrivateScope VarScope(*this); |
| 2236 | VarScope.addPrivate(OrigVD, [OrigAddr]() { return OrigAddr; }); |
| 2237 | (void)VarScope.Privatize(); |
| 2238 | EmitIgnoredExpr(F); |
| 2239 | } |
| 2240 | ++IC; |
| 2241 | ++IPC; |
| 2242 | } |
| 2243 | if (DoneBB) |
| 2244 | EmitBlock(DoneBB, /*IsFinished=*/true); |
| 2245 | } |
| 2246 | |
| 2247 | static void emitOMPLoopBodyWithStopPoint(CodeGenFunction &CGF, |
| 2248 | const OMPLoopDirective &S, |
| 2249 | CodeGenFunction::JumpDest LoopExit) { |
| 2250 | CGF.EmitOMPLoopBody(S, LoopExit); |
| 2251 | CGF.EmitStopPoint(&S); |
| 2252 | } |
| 2253 | |
| 2254 | /// Emit a helper variable and return corresponding lvalue. |
| 2255 | static LValue EmitOMPHelperVar(CodeGenFunction &CGF, |
| 2256 | const DeclRefExpr *Helper) { |
| 2257 | auto VDecl = cast<VarDecl>(Helper->getDecl()); |
| 2258 | CGF.EmitVarDecl(*VDecl); |
| 2259 | return CGF.EmitLValue(Helper); |
| 2260 | } |
| 2261 | |
| 2262 | static void emitCommonSimdLoop(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2263 | const RegionCodeGenTy &SimdInitGen, |
| 2264 | const RegionCodeGenTy &BodyCodeGen) { |
| 2265 | auto &&ThenGen = [&S, &SimdInitGen, &BodyCodeGen](CodeGenFunction &CGF, |
| 2266 | PrePostActionTy &) { |
| 2267 | CGOpenMPRuntime::NontemporalDeclsRAII NontemporalsRegion(CGF.CGM, S); |
| 2268 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 2269 | SimdInitGen(CGF); |
| 2270 | |
| 2271 | BodyCodeGen(CGF); |
| 2272 | }; |
| 2273 | auto &&ElseGen = [&BodyCodeGen](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2274 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 2275 | CGF.LoopStack.setVectorizeEnable(/*Enable=*/false); |
| 2276 | |
| 2277 | BodyCodeGen(CGF); |
| 2278 | }; |
| 2279 | const Expr *IfCond = nullptr; |
| 2280 | if (isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 2281 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 2282 | if (CGF.getLangOpts().OpenMP >= 50 && |
| 2283 | (C->getNameModifier() == OMPD_unknown || |
| 2284 | C->getNameModifier() == OMPD_simd)) { |
| 2285 | IfCond = C->getCondition(); |
| 2286 | break; |
| 2287 | } |
| 2288 | } |
| 2289 | } |
| 2290 | if (IfCond) { |
| 2291 | CGF.CGM.getOpenMPRuntime().emitIfClause(CGF, IfCond, ThenGen, ElseGen); |
| 2292 | } else { |
| 2293 | RegionCodeGenTy ThenRCG(ThenGen); |
| 2294 | ThenRCG(CGF); |
| 2295 | } |
| 2296 | } |
| 2297 | |
| 2298 | static void emitOMPSimdRegion(CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 2299 | PrePostActionTy &Action) { |
| 2300 | Action.Enter(CGF); |
| 2301 | assert(isOpenMPSimdDirective(S.getDirectiveKind()) && |
| 2302 | "Expected simd directive" ); |
| 2303 | OMPLoopScope PreInitScope(CGF, S); |
| 2304 | // if (PreCond) { |
| 2305 | // for (IV in 0..LastIteration) BODY; |
| 2306 | // <Final counter/linear vars updates>; |
| 2307 | // } |
| 2308 | // |
| 2309 | if (isOpenMPDistributeDirective(S.getDirectiveKind()) || |
| 2310 | isOpenMPWorksharingDirective(S.getDirectiveKind()) || |
| 2311 | isOpenMPTaskLoopDirective(S.getDirectiveKind())) { |
| 2312 | (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable())); |
| 2313 | (void)EmitOMPHelperVar(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable())); |
| 2314 | } |
| 2315 | |
| 2316 | // Emit: if (PreCond) - begin. |
| 2317 | // If the condition constant folds and can be elided, avoid emitting the |
| 2318 | // whole loop. |
| 2319 | bool CondConstant; |
| 2320 | llvm::BasicBlock *ContBlock = nullptr; |
| 2321 | if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) { |
| 2322 | if (!CondConstant) |
| 2323 | return; |
| 2324 | } else { |
| 2325 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("simd.if.then" ); |
| 2326 | ContBlock = CGF.createBasicBlock("simd.if.end" ); |
| 2327 | emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock, |
| 2328 | CGF.getProfileCount(&S)); |
| 2329 | CGF.EmitBlock(ThenBlock); |
| 2330 | CGF.incrementProfileCounter(&S); |
| 2331 | } |
| 2332 | |
| 2333 | // Emit the loop iteration variable. |
| 2334 | const Expr *IVExpr = S.getIterationVariable(); |
| 2335 | const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl()); |
| 2336 | CGF.EmitVarDecl(*IVDecl); |
| 2337 | CGF.EmitIgnoredExpr(S.getInit()); |
| 2338 | |
| 2339 | // Emit the iterations count variable. |
| 2340 | // If it is not a variable, Sema decided to calculate iterations count on |
| 2341 | // each iteration (e.g., it is foldable into a constant). |
| 2342 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) { |
| 2343 | CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl())); |
| 2344 | // Emit calculation of the iterations count. |
| 2345 | CGF.EmitIgnoredExpr(S.getCalcLastIteration()); |
| 2346 | } |
| 2347 | |
| 2348 | emitAlignedClause(CGF, S); |
| 2349 | (void)CGF.EmitOMPLinearClauseInit(S); |
| 2350 | { |
| 2351 | CodeGenFunction::OMPPrivateScope LoopScope(CGF); |
| 2352 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope); |
| 2353 | CGF.EmitOMPLinearClause(S, LoopScope); |
| 2354 | CGF.EmitOMPPrivateClause(S, LoopScope); |
| 2355 | CGF.EmitOMPReductionClauseInit(S, LoopScope); |
| 2356 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion( |
| 2357 | CGF, S, CGF.EmitLValue(S.getIterationVariable())); |
| 2358 | bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope); |
| 2359 | (void)LoopScope.Privatize(); |
| 2360 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 2361 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S); |
| 2362 | |
| 2363 | emitCommonSimdLoop( |
| 2364 | CGF, S, |
| 2365 | [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2366 | CGF.EmitOMPSimdInit(S); |
| 2367 | }, |
| 2368 | [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2369 | CGF.EmitOMPInnerLoop( |
| 2370 | S, LoopScope.requiresCleanups(), S.getCond(), S.getInc(), |
| 2371 | [&S](CodeGenFunction &CGF) { |
| 2372 | emitOMPLoopBodyWithStopPoint(CGF, S, |
| 2373 | CodeGenFunction::JumpDest()); |
| 2374 | }, |
| 2375 | [](CodeGenFunction &) {}); |
| 2376 | }); |
| 2377 | CGF.EmitOMPSimdFinal(S, [](CodeGenFunction &) { return nullptr; }); |
| 2378 | // Emit final copy of the lastprivate variables at the end of loops. |
| 2379 | if (HasLastprivateClause) |
| 2380 | CGF.EmitOMPLastprivateClauseFinal(S, /*NoFinals=*/true); |
| 2381 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_simd); |
| 2382 | emitPostUpdateForReductionClause(CGF, S, |
| 2383 | [](CodeGenFunction &) { return nullptr; }); |
| 2384 | } |
| 2385 | CGF.EmitOMPLinearClauseFinal(S, [](CodeGenFunction &) { return nullptr; }); |
| 2386 | // Emit: if (PreCond) - end. |
| 2387 | if (ContBlock) { |
| 2388 | CGF.EmitBranch(ContBlock); |
| 2389 | CGF.EmitBlock(ContBlock, true); |
| 2390 | } |
| 2391 | } |
| 2392 | |
| 2393 | void CodeGenFunction::EmitOMPSimdDirective(const OMPSimdDirective &S) { |
| 2394 | ParentLoopDirectiveForScanRegion ScanRegion(*this, S); |
| 2395 | OMPFirstScanLoop = true; |
| 2396 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 2397 | emitOMPSimdRegion(CGF, S, Action); |
| 2398 | }; |
| 2399 | { |
| 2400 | auto LPCRegion = |
| 2401 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 2402 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 2403 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen); |
| 2404 | } |
| 2405 | // Check for outer lastprivate conditional update. |
| 2406 | checkForLastprivateConditionalUpdate(*this, S); |
| 2407 | } |
| 2408 | |
| 2409 | void CodeGenFunction::EmitOMPOuterLoop( |
| 2410 | bool DynamicOrOrdered, bool IsMonotonic, const OMPLoopDirective &S, |
| 2411 | CodeGenFunction::OMPPrivateScope &LoopScope, |
| 2412 | const CodeGenFunction::OMPLoopArguments &LoopArgs, |
| 2413 | const CodeGenFunction::CodeGenLoopTy &CodeGenLoop, |
| 2414 | const CodeGenFunction::CodeGenOrderedTy &CodeGenOrdered) { |
| 2415 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 2416 | |
| 2417 | const Expr *IVExpr = S.getIterationVariable(); |
| 2418 | const unsigned IVSize = getContext().getTypeSize(IVExpr->getType()); |
| 2419 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 2420 | |
| 2421 | JumpDest LoopExit = getJumpDestInCurrentScope("omp.dispatch.end" ); |
| 2422 | |
| 2423 | // Start the loop with a block that tests the condition. |
| 2424 | llvm::BasicBlock *CondBlock = createBasicBlock("omp.dispatch.cond" ); |
| 2425 | EmitBlock(CondBlock); |
| 2426 | const SourceRange R = S.getSourceRange(); |
| 2427 | LoopStack.push(CondBlock, SourceLocToDebugLoc(R.getBegin()), |
| 2428 | SourceLocToDebugLoc(R.getEnd())); |
| 2429 | |
| 2430 | llvm::Value *BoolCondVal = nullptr; |
| 2431 | if (!DynamicOrOrdered) { |
| 2432 | // UB = min(UB, GlobalUB) or |
| 2433 | // UB = min(UB, PrevUB) for combined loop sharing constructs (e.g. |
| 2434 | // 'distribute parallel for') |
| 2435 | EmitIgnoredExpr(LoopArgs.EUB); |
| 2436 | // IV = LB |
| 2437 | EmitIgnoredExpr(LoopArgs.Init); |
| 2438 | // IV < UB |
| 2439 | BoolCondVal = EvaluateExprAsBool(LoopArgs.Cond); |
| 2440 | } else { |
| 2441 | BoolCondVal = |
| 2442 | RT.emitForNext(*this, S.getBeginLoc(), IVSize, IVSigned, LoopArgs.IL, |
| 2443 | LoopArgs.LB, LoopArgs.UB, LoopArgs.ST); |
| 2444 | } |
| 2445 | |
| 2446 | // If there are any cleanups between here and the loop-exit scope, |
| 2447 | // create a block to stage a loop exit along. |
| 2448 | llvm::BasicBlock *ExitBlock = LoopExit.getBlock(); |
| 2449 | if (LoopScope.requiresCleanups()) |
| 2450 | ExitBlock = createBasicBlock("omp.dispatch.cleanup" ); |
| 2451 | |
| 2452 | llvm::BasicBlock *LoopBody = createBasicBlock("omp.dispatch.body" ); |
| 2453 | Builder.CreateCondBr(BoolCondVal, LoopBody, ExitBlock); |
| 2454 | if (ExitBlock != LoopExit.getBlock()) { |
| 2455 | EmitBlock(ExitBlock); |
| 2456 | EmitBranchThroughCleanup(LoopExit); |
| 2457 | } |
| 2458 | EmitBlock(LoopBody); |
| 2459 | |
| 2460 | // Emit "IV = LB" (in case of static schedule, we have already calculated new |
| 2461 | // LB for loop condition and emitted it above). |
| 2462 | if (DynamicOrOrdered) |
| 2463 | EmitIgnoredExpr(LoopArgs.Init); |
| 2464 | |
| 2465 | // Create a block for the increment. |
| 2466 | JumpDest Continue = getJumpDestInCurrentScope("omp.dispatch.inc" ); |
| 2467 | BreakContinueStack.push_back(BreakContinue(LoopExit, Continue)); |
| 2468 | |
| 2469 | emitCommonSimdLoop( |
| 2470 | *this, S, |
| 2471 | [&S, IsMonotonic](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2472 | // Generate !llvm.loop.parallel metadata for loads and stores for loops |
| 2473 | // with dynamic/guided scheduling and without ordered clause. |
| 2474 | if (!isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 2475 | CGF.LoopStack.setParallel(!IsMonotonic); |
| 2476 | if (const auto *C = S.getSingleClause<OMPOrderClause>()) |
| 2477 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 2478 | CGF.LoopStack.setParallel(/*Enable=*/true); |
| 2479 | } else { |
| 2480 | CGF.EmitOMPSimdInit(S, IsMonotonic); |
| 2481 | } |
| 2482 | }, |
| 2483 | [&S, &LoopArgs, LoopExit, &CodeGenLoop, IVSize, IVSigned, &CodeGenOrdered, |
| 2484 | &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2485 | SourceLocation Loc = S.getBeginLoc(); |
| 2486 | // when 'distribute' is not combined with a 'for': |
| 2487 | // while (idx <= UB) { BODY; ++idx; } |
| 2488 | // when 'distribute' is combined with a 'for' |
| 2489 | // (e.g. 'distribute parallel for') |
| 2490 | // while (idx <= UB) { <CodeGen rest of pragma>; idx += ST; } |
| 2491 | CGF.EmitOMPInnerLoop( |
| 2492 | S, LoopScope.requiresCleanups(), LoopArgs.Cond, LoopArgs.IncExpr, |
| 2493 | [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) { |
| 2494 | CodeGenLoop(CGF, S, LoopExit); |
| 2495 | }, |
| 2496 | [IVSize, IVSigned, Loc, &CodeGenOrdered](CodeGenFunction &CGF) { |
| 2497 | CodeGenOrdered(CGF, Loc, IVSize, IVSigned); |
| 2498 | }); |
| 2499 | }); |
| 2500 | |
| 2501 | EmitBlock(Continue.getBlock()); |
| 2502 | BreakContinueStack.pop_back(); |
| 2503 | if (!DynamicOrOrdered) { |
| 2504 | // Emit "LB = LB + Stride", "UB = UB + Stride". |
| 2505 | EmitIgnoredExpr(LoopArgs.NextLB); |
| 2506 | EmitIgnoredExpr(LoopArgs.NextUB); |
| 2507 | } |
| 2508 | |
| 2509 | EmitBranch(CondBlock); |
| 2510 | LoopStack.pop(); |
| 2511 | // Emit the fall-through block. |
| 2512 | EmitBlock(LoopExit.getBlock()); |
| 2513 | |
| 2514 | // Tell the runtime we are done. |
| 2515 | auto &&CodeGen = [DynamicOrOrdered, &S](CodeGenFunction &CGF) { |
| 2516 | if (!DynamicOrOrdered) |
| 2517 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(), |
| 2518 | S.getDirectiveKind()); |
| 2519 | }; |
| 2520 | OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen); |
| 2521 | } |
| 2522 | |
| 2523 | void CodeGenFunction::EmitOMPForOuterLoop( |
| 2524 | const OpenMPScheduleTy &ScheduleKind, bool IsMonotonic, |
| 2525 | const OMPLoopDirective &S, OMPPrivateScope &LoopScope, bool Ordered, |
| 2526 | const OMPLoopArguments &LoopArgs, |
| 2527 | const CodeGenDispatchBoundsTy &CGDispatchBounds) { |
| 2528 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 2529 | |
| 2530 | // Dynamic scheduling of the outer loop (dynamic, guided, auto, runtime). |
| 2531 | const bool DynamicOrOrdered = |
| 2532 | Ordered || RT.isDynamic(ScheduleKind.Schedule); |
| 2533 | |
| 2534 | assert((Ordered || |
| 2535 | !RT.isStaticNonchunked(ScheduleKind.Schedule, |
| 2536 | LoopArgs.Chunk != nullptr)) && |
| 2537 | "static non-chunked schedule does not need outer loop" ); |
| 2538 | |
| 2539 | // Emit outer loop. |
| 2540 | // |
| 2541 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 2542 | // When schedule(dynamic,chunk_size) is specified, the iterations are |
| 2543 | // distributed to threads in the team in chunks as the threads request them. |
| 2544 | // Each thread executes a chunk of iterations, then requests another chunk, |
| 2545 | // until no chunks remain to be distributed. Each chunk contains chunk_size |
| 2546 | // iterations, except for the last chunk to be distributed, which may have |
| 2547 | // fewer iterations. When no chunk_size is specified, it defaults to 1. |
| 2548 | // |
| 2549 | // When schedule(guided,chunk_size) is specified, the iterations are assigned |
| 2550 | // to threads in the team in chunks as the executing threads request them. |
| 2551 | // Each thread executes a chunk of iterations, then requests another chunk, |
| 2552 | // until no chunks remain to be assigned. For a chunk_size of 1, the size of |
| 2553 | // each chunk is proportional to the number of unassigned iterations divided |
| 2554 | // by the number of threads in the team, decreasing to 1. For a chunk_size |
| 2555 | // with value k (greater than 1), the size of each chunk is determined in the |
| 2556 | // same way, with the restriction that the chunks do not contain fewer than k |
| 2557 | // iterations (except for the last chunk to be assigned, which may have fewer |
| 2558 | // than k iterations). |
| 2559 | // |
| 2560 | // When schedule(auto) is specified, the decision regarding scheduling is |
| 2561 | // delegated to the compiler and/or runtime system. The programmer gives the |
| 2562 | // implementation the freedom to choose any possible mapping of iterations to |
| 2563 | // threads in the team. |
| 2564 | // |
| 2565 | // When schedule(runtime) is specified, the decision regarding scheduling is |
| 2566 | // deferred until run time, and the schedule and chunk size are taken from the |
| 2567 | // run-sched-var ICV. If the ICV is set to auto, the schedule is |
| 2568 | // implementation defined |
| 2569 | // |
| 2570 | // while(__kmpc_dispatch_next(&LB, &UB)) { |
| 2571 | // idx = LB; |
| 2572 | // while (idx <= UB) { BODY; ++idx; |
| 2573 | // __kmpc_dispatch_fini_(4|8)[u](); // For ordered loops only. |
| 2574 | // } // inner loop |
| 2575 | // } |
| 2576 | // |
| 2577 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 2578 | // When schedule(static, chunk_size) is specified, iterations are divided into |
| 2579 | // chunks of size chunk_size, and the chunks are assigned to the threads in |
| 2580 | // the team in a round-robin fashion in the order of the thread number. |
| 2581 | // |
| 2582 | // while(UB = min(UB, GlobalUB), idx = LB, idx < UB) { |
| 2583 | // while (idx <= UB) { BODY; ++idx; } // inner loop |
| 2584 | // LB = LB + ST; |
| 2585 | // UB = UB + ST; |
| 2586 | // } |
| 2587 | // |
| 2588 | |
| 2589 | const Expr *IVExpr = S.getIterationVariable(); |
| 2590 | const unsigned IVSize = getContext().getTypeSize(IVExpr->getType()); |
| 2591 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 2592 | |
| 2593 | if (DynamicOrOrdered) { |
| 2594 | const std::pair<llvm::Value *, llvm::Value *> DispatchBounds = |
| 2595 | CGDispatchBounds(*this, S, LoopArgs.LB, LoopArgs.UB); |
| 2596 | llvm::Value *LBVal = DispatchBounds.first; |
| 2597 | llvm::Value *UBVal = DispatchBounds.second; |
| 2598 | CGOpenMPRuntime::DispatchRTInput DipatchRTInputValues = {LBVal, UBVal, |
| 2599 | LoopArgs.Chunk}; |
| 2600 | RT.emitForDispatchInit(*this, S.getBeginLoc(), ScheduleKind, IVSize, |
| 2601 | IVSigned, Ordered, DipatchRTInputValues); |
| 2602 | } else { |
| 2603 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 2604 | IVSize, IVSigned, Ordered, LoopArgs.IL, LoopArgs.LB, LoopArgs.UB, |
| 2605 | LoopArgs.ST, LoopArgs.Chunk); |
| 2606 | RT.emitForStaticInit(*this, S.getBeginLoc(), S.getDirectiveKind(), |
| 2607 | ScheduleKind, StaticInit); |
| 2608 | } |
| 2609 | |
| 2610 | auto &&CodeGenOrdered = [Ordered](CodeGenFunction &CGF, SourceLocation Loc, |
| 2611 | const unsigned IVSize, |
| 2612 | const bool IVSigned) { |
| 2613 | if (Ordered) { |
| 2614 | CGF.CGM.getOpenMPRuntime().emitForOrderedIterationEnd(CGF, Loc, IVSize, |
| 2615 | IVSigned); |
| 2616 | } |
| 2617 | }; |
| 2618 | |
| 2619 | OMPLoopArguments OuterLoopArgs(LoopArgs.LB, LoopArgs.UB, LoopArgs.ST, |
| 2620 | LoopArgs.IL, LoopArgs.Chunk, LoopArgs.EUB); |
| 2621 | OuterLoopArgs.IncExpr = S.getInc(); |
| 2622 | OuterLoopArgs.Init = S.getInit(); |
| 2623 | OuterLoopArgs.Cond = S.getCond(); |
| 2624 | OuterLoopArgs.NextLB = S.getNextLowerBound(); |
| 2625 | OuterLoopArgs.NextUB = S.getNextUpperBound(); |
| 2626 | EmitOMPOuterLoop(DynamicOrOrdered, IsMonotonic, S, LoopScope, OuterLoopArgs, |
| 2627 | emitOMPLoopBodyWithStopPoint, CodeGenOrdered); |
| 2628 | } |
| 2629 | |
| 2630 | static void emitEmptyOrdered(CodeGenFunction &, SourceLocation Loc, |
| 2631 | const unsigned IVSize, const bool IVSigned) {} |
| 2632 | |
| 2633 | void CodeGenFunction::EmitOMPDistributeOuterLoop( |
| 2634 | OpenMPDistScheduleClauseKind ScheduleKind, const OMPLoopDirective &S, |
| 2635 | OMPPrivateScope &LoopScope, const OMPLoopArguments &LoopArgs, |
| 2636 | const CodeGenLoopTy &CodeGenLoopContent) { |
| 2637 | |
| 2638 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 2639 | |
| 2640 | // Emit outer loop. |
| 2641 | // Same behavior as a OMPForOuterLoop, except that schedule cannot be |
| 2642 | // dynamic |
| 2643 | // |
| 2644 | |
| 2645 | const Expr *IVExpr = S.getIterationVariable(); |
| 2646 | const unsigned IVSize = getContext().getTypeSize(IVExpr->getType()); |
| 2647 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 2648 | |
| 2649 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 2650 | IVSize, IVSigned, /* Ordered = */ false, LoopArgs.IL, LoopArgs.LB, |
| 2651 | LoopArgs.UB, LoopArgs.ST, LoopArgs.Chunk); |
| 2652 | RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind, StaticInit); |
| 2653 | |
| 2654 | // for combined 'distribute' and 'for' the increment expression of distribute |
| 2655 | // is stored in DistInc. For 'distribute' alone, it is in Inc. |
| 2656 | Expr *IncExpr; |
| 2657 | if (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind())) |
| 2658 | IncExpr = S.getDistInc(); |
| 2659 | else |
| 2660 | IncExpr = S.getInc(); |
| 2661 | |
| 2662 | // this routine is shared by 'omp distribute parallel for' and |
| 2663 | // 'omp distribute': select the right EUB expression depending on the |
| 2664 | // directive |
| 2665 | OMPLoopArguments OuterLoopArgs; |
| 2666 | OuterLoopArgs.LB = LoopArgs.LB; |
| 2667 | OuterLoopArgs.UB = LoopArgs.UB; |
| 2668 | OuterLoopArgs.ST = LoopArgs.ST; |
| 2669 | OuterLoopArgs.IL = LoopArgs.IL; |
| 2670 | OuterLoopArgs.Chunk = LoopArgs.Chunk; |
| 2671 | OuterLoopArgs.EUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 2672 | ? S.getCombinedEnsureUpperBound() |
| 2673 | : S.getEnsureUpperBound(); |
| 2674 | OuterLoopArgs.IncExpr = IncExpr; |
| 2675 | OuterLoopArgs.Init = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 2676 | ? S.getCombinedInit() |
| 2677 | : S.getInit(); |
| 2678 | OuterLoopArgs.Cond = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 2679 | ? S.getCombinedCond() |
| 2680 | : S.getCond(); |
| 2681 | OuterLoopArgs.NextLB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 2682 | ? S.getCombinedNextLowerBound() |
| 2683 | : S.getNextLowerBound(); |
| 2684 | OuterLoopArgs.NextUB = isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 2685 | ? S.getCombinedNextUpperBound() |
| 2686 | : S.getNextUpperBound(); |
| 2687 | |
| 2688 | EmitOMPOuterLoop(/* DynamicOrOrdered = */ false, /* IsMonotonic = */ false, S, |
| 2689 | LoopScope, OuterLoopArgs, CodeGenLoopContent, |
| 2690 | emitEmptyOrdered); |
| 2691 | } |
| 2692 | |
| 2693 | static std::pair<LValue, LValue> |
| 2694 | emitDistributeParallelForInnerBounds(CodeGenFunction &CGF, |
| 2695 | const OMPExecutableDirective &S) { |
| 2696 | const OMPLoopDirective &LS = cast<OMPLoopDirective>(S); |
| 2697 | LValue LB = |
| 2698 | EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable())); |
| 2699 | LValue UB = |
| 2700 | EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable())); |
| 2701 | |
| 2702 | // When composing 'distribute' with 'for' (e.g. as in 'distribute |
| 2703 | // parallel for') we need to use the 'distribute' |
| 2704 | // chunk lower and upper bounds rather than the whole loop iteration |
| 2705 | // space. These are parameters to the outlined function for 'parallel' |
| 2706 | // and we copy the bounds of the previous schedule into the |
| 2707 | // the current ones. |
| 2708 | LValue PrevLB = CGF.EmitLValue(LS.getPrevLowerBoundVariable()); |
| 2709 | LValue PrevUB = CGF.EmitLValue(LS.getPrevUpperBoundVariable()); |
| 2710 | llvm::Value *PrevLBVal = CGF.EmitLoadOfScalar( |
| 2711 | PrevLB, LS.getPrevLowerBoundVariable()->getExprLoc()); |
| 2712 | PrevLBVal = CGF.EmitScalarConversion( |
| 2713 | PrevLBVal, LS.getPrevLowerBoundVariable()->getType(), |
| 2714 | LS.getIterationVariable()->getType(), |
| 2715 | LS.getPrevLowerBoundVariable()->getExprLoc()); |
| 2716 | llvm::Value *PrevUBVal = CGF.EmitLoadOfScalar( |
| 2717 | PrevUB, LS.getPrevUpperBoundVariable()->getExprLoc()); |
| 2718 | PrevUBVal = CGF.EmitScalarConversion( |
| 2719 | PrevUBVal, LS.getPrevUpperBoundVariable()->getType(), |
| 2720 | LS.getIterationVariable()->getType(), |
| 2721 | LS.getPrevUpperBoundVariable()->getExprLoc()); |
| 2722 | |
| 2723 | CGF.EmitStoreOfScalar(PrevLBVal, LB); |
| 2724 | CGF.EmitStoreOfScalar(PrevUBVal, UB); |
| 2725 | |
| 2726 | return {LB, UB}; |
| 2727 | } |
| 2728 | |
| 2729 | /// if the 'for' loop has a dispatch schedule (e.g. dynamic, guided) then |
| 2730 | /// we need to use the LB and UB expressions generated by the worksharing |
| 2731 | /// code generation support, whereas in non combined situations we would |
| 2732 | /// just emit 0 and the LastIteration expression |
| 2733 | /// This function is necessary due to the difference of the LB and UB |
| 2734 | /// types for the RT emission routines for 'for_static_init' and |
| 2735 | /// 'for_dispatch_init' |
| 2736 | static std::pair<llvm::Value *, llvm::Value *> |
| 2737 | emitDistributeParallelForDispatchBounds(CodeGenFunction &CGF, |
| 2738 | const OMPExecutableDirective &S, |
| 2739 | Address LB, Address UB) { |
| 2740 | const OMPLoopDirective &LS = cast<OMPLoopDirective>(S); |
| 2741 | const Expr *IVExpr = LS.getIterationVariable(); |
| 2742 | // when implementing a dynamic schedule for a 'for' combined with a |
| 2743 | // 'distribute' (e.g. 'distribute parallel for'), the 'for' loop |
| 2744 | // is not normalized as each team only executes its own assigned |
| 2745 | // distribute chunk |
| 2746 | QualType IteratorTy = IVExpr->getType(); |
| 2747 | llvm::Value *LBVal = |
| 2748 | CGF.EmitLoadOfScalar(LB, /*Volatile=*/false, IteratorTy, S.getBeginLoc()); |
| 2749 | llvm::Value *UBVal = |
| 2750 | CGF.EmitLoadOfScalar(UB, /*Volatile=*/false, IteratorTy, S.getBeginLoc()); |
| 2751 | return {LBVal, UBVal}; |
| 2752 | } |
| 2753 | |
| 2754 | static void emitDistributeParallelForDistributeInnerBoundParams( |
| 2755 | CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 2756 | llvm::SmallVectorImpl<llvm::Value *> &CapturedVars) { |
| 2757 | const auto &Dir = cast<OMPLoopDirective>(S); |
| 2758 | LValue LB = |
| 2759 | CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedLowerBoundVariable())); |
| 2760 | llvm::Value *LBCast = |
| 2761 | CGF.Builder.CreateIntCast(CGF.Builder.CreateLoad(LB.getAddress(CGF)), |
| 2762 | CGF.SizeTy, /*isSigned=*/false); |
| 2763 | CapturedVars.push_back(LBCast); |
| 2764 | LValue UB = |
| 2765 | CGF.EmitLValue(cast<DeclRefExpr>(Dir.getCombinedUpperBoundVariable())); |
| 2766 | |
| 2767 | llvm::Value *UBCast = |
| 2768 | CGF.Builder.CreateIntCast(CGF.Builder.CreateLoad(UB.getAddress(CGF)), |
| 2769 | CGF.SizeTy, /*isSigned=*/false); |
| 2770 | CapturedVars.push_back(UBCast); |
| 2771 | } |
| 2772 | |
| 2773 | static void |
| 2774 | emitInnerParallelForWhenCombined(CodeGenFunction &CGF, |
| 2775 | const OMPLoopDirective &S, |
| 2776 | CodeGenFunction::JumpDest LoopExit) { |
| 2777 | auto &&CGInlinedWorksharingLoop = [&S](CodeGenFunction &CGF, |
| 2778 | PrePostActionTy &Action) { |
| 2779 | Action.Enter(CGF); |
| 2780 | bool HasCancel = false; |
| 2781 | if (!isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 2782 | if (const auto *D = dyn_cast<OMPTeamsDistributeParallelForDirective>(&S)) |
| 2783 | HasCancel = D->hasCancel(); |
| 2784 | else if (const auto *D = dyn_cast<OMPDistributeParallelForDirective>(&S)) |
| 2785 | HasCancel = D->hasCancel(); |
| 2786 | else if (const auto *D = |
| 2787 | dyn_cast<OMPTargetTeamsDistributeParallelForDirective>(&S)) |
| 2788 | HasCancel = D->hasCancel(); |
| 2789 | } |
| 2790 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, S.getDirectiveKind(), |
| 2791 | HasCancel); |
| 2792 | CGF.EmitOMPWorksharingLoop(S, S.getPrevEnsureUpperBound(), |
| 2793 | emitDistributeParallelForInnerBounds, |
| 2794 | emitDistributeParallelForDispatchBounds); |
| 2795 | }; |
| 2796 | |
| 2797 | emitCommonOMPParallelDirective( |
| 2798 | CGF, S, |
| 2799 | isOpenMPSimdDirective(S.getDirectiveKind()) ? OMPD_for_simd : OMPD_for, |
| 2800 | CGInlinedWorksharingLoop, |
| 2801 | emitDistributeParallelForDistributeInnerBoundParams); |
| 2802 | } |
| 2803 | |
| 2804 | void CodeGenFunction::EmitOMPDistributeParallelForDirective( |
| 2805 | const OMPDistributeParallelForDirective &S) { |
| 2806 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2807 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 2808 | S.getDistInc()); |
| 2809 | }; |
| 2810 | OMPLexicalScope Scope(*this, S, OMPD_parallel); |
| 2811 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen); |
| 2812 | } |
| 2813 | |
| 2814 | void CodeGenFunction::EmitOMPDistributeParallelForSimdDirective( |
| 2815 | const OMPDistributeParallelForSimdDirective &S) { |
| 2816 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2817 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 2818 | S.getDistInc()); |
| 2819 | }; |
| 2820 | OMPLexicalScope Scope(*this, S, OMPD_parallel); |
| 2821 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen); |
| 2822 | } |
| 2823 | |
| 2824 | void CodeGenFunction::EmitOMPDistributeSimdDirective( |
| 2825 | const OMPDistributeSimdDirective &S) { |
| 2826 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 2827 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 2828 | }; |
| 2829 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 2830 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen); |
| 2831 | } |
| 2832 | |
| 2833 | void CodeGenFunction::EmitOMPTargetSimdDeviceFunction( |
| 2834 | CodeGenModule &CGM, StringRef ParentName, const OMPTargetSimdDirective &S) { |
| 2835 | // Emit SPMD target parallel for region as a standalone region. |
| 2836 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 2837 | emitOMPSimdRegion(CGF, S, Action); |
| 2838 | }; |
| 2839 | llvm::Function *Fn; |
| 2840 | llvm::Constant *Addr; |
| 2841 | // Emit target region as a standalone region. |
| 2842 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 2843 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 2844 | assert(Fn && Addr && "Target device function emission failed." ); |
| 2845 | } |
| 2846 | |
| 2847 | void CodeGenFunction::EmitOMPTargetSimdDirective( |
| 2848 | const OMPTargetSimdDirective &S) { |
| 2849 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 2850 | emitOMPSimdRegion(CGF, S, Action); |
| 2851 | }; |
| 2852 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 2853 | } |
| 2854 | |
| 2855 | namespace { |
| 2856 | struct ScheduleKindModifiersTy { |
| 2857 | OpenMPScheduleClauseKind Kind; |
| 2858 | OpenMPScheduleClauseModifier M1; |
| 2859 | OpenMPScheduleClauseModifier M2; |
| 2860 | ScheduleKindModifiersTy(OpenMPScheduleClauseKind Kind, |
| 2861 | OpenMPScheduleClauseModifier M1, |
| 2862 | OpenMPScheduleClauseModifier M2) |
| 2863 | : Kind(Kind), M1(M1), M2(M2) {} |
| 2864 | }; |
| 2865 | } // namespace |
| 2866 | |
| 2867 | bool CodeGenFunction::EmitOMPWorksharingLoop( |
| 2868 | const OMPLoopDirective &S, Expr *EUB, |
| 2869 | const CodeGenLoopBoundsTy &CodeGenLoopBounds, |
| 2870 | const CodeGenDispatchBoundsTy &CGDispatchBounds) { |
| 2871 | // Emit the loop iteration variable. |
| 2872 | const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable()); |
| 2873 | const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl()); |
| 2874 | EmitVarDecl(*IVDecl); |
| 2875 | |
| 2876 | // Emit the iterations count variable. |
| 2877 | // If it is not a variable, Sema decided to calculate iterations count on each |
| 2878 | // iteration (e.g., it is foldable into a constant). |
| 2879 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) { |
| 2880 | EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl())); |
| 2881 | // Emit calculation of the iterations count. |
| 2882 | EmitIgnoredExpr(S.getCalcLastIteration()); |
| 2883 | } |
| 2884 | |
| 2885 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 2886 | |
| 2887 | bool HasLastprivateClause; |
| 2888 | // Check pre-condition. |
| 2889 | { |
| 2890 | OMPLoopScope PreInitScope(*this, S); |
| 2891 | // Skip the entire loop if we don't meet the precondition. |
| 2892 | // If the condition constant folds and can be elided, avoid emitting the |
| 2893 | // whole loop. |
| 2894 | bool CondConstant; |
| 2895 | llvm::BasicBlock *ContBlock = nullptr; |
| 2896 | if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) { |
| 2897 | if (!CondConstant) |
| 2898 | return false; |
| 2899 | } else { |
| 2900 | llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then" ); |
| 2901 | ContBlock = createBasicBlock("omp.precond.end" ); |
| 2902 | emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock, |
| 2903 | getProfileCount(&S)); |
| 2904 | EmitBlock(ThenBlock); |
| 2905 | incrementProfileCounter(&S); |
| 2906 | } |
| 2907 | |
| 2908 | RunCleanupsScope DoacrossCleanupScope(*this); |
| 2909 | bool Ordered = false; |
| 2910 | if (const auto *OrderedClause = S.getSingleClause<OMPOrderedClause>()) { |
| 2911 | if (OrderedClause->getNumForLoops()) |
| 2912 | RT.emitDoacrossInit(*this, S, OrderedClause->getLoopNumIterations()); |
| 2913 | else |
| 2914 | Ordered = true; |
| 2915 | } |
| 2916 | |
| 2917 | llvm::DenseSet<const Expr *> EmittedFinals; |
| 2918 | emitAlignedClause(*this, S); |
| 2919 | bool HasLinears = EmitOMPLinearClauseInit(S); |
| 2920 | // Emit helper vars inits. |
| 2921 | |
| 2922 | std::pair<LValue, LValue> Bounds = CodeGenLoopBounds(*this, S); |
| 2923 | LValue LB = Bounds.first; |
| 2924 | LValue UB = Bounds.second; |
| 2925 | LValue ST = |
| 2926 | EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable())); |
| 2927 | LValue IL = |
| 2928 | EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable())); |
| 2929 | |
| 2930 | // Emit 'then' code. |
| 2931 | { |
| 2932 | OMPPrivateScope LoopScope(*this); |
| 2933 | if (EmitOMPFirstprivateClause(S, LoopScope) || HasLinears) { |
| 2934 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 2935 | // initialization of firstprivate variables and post-update of |
| 2936 | // lastprivate variables. |
| 2937 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 2938 | *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false, |
| 2939 | /*ForceSimpleCall=*/true); |
| 2940 | } |
| 2941 | EmitOMPPrivateClause(S, LoopScope); |
| 2942 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion( |
| 2943 | *this, S, EmitLValue(S.getIterationVariable())); |
| 2944 | HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope); |
| 2945 | EmitOMPReductionClauseInit(S, LoopScope); |
| 2946 | EmitOMPPrivateLoopCounters(S, LoopScope); |
| 2947 | EmitOMPLinearClause(S, LoopScope); |
| 2948 | (void)LoopScope.Privatize(); |
| 2949 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 2950 | CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S); |
| 2951 | |
| 2952 | // Detect the loop schedule kind and chunk. |
| 2953 | const Expr *ChunkExpr = nullptr; |
| 2954 | OpenMPScheduleTy ScheduleKind; |
| 2955 | if (const auto *C = S.getSingleClause<OMPScheduleClause>()) { |
| 2956 | ScheduleKind.Schedule = C->getScheduleKind(); |
| 2957 | ScheduleKind.M1 = C->getFirstScheduleModifier(); |
| 2958 | ScheduleKind.M2 = C->getSecondScheduleModifier(); |
| 2959 | ChunkExpr = C->getChunkSize(); |
| 2960 | } else { |
| 2961 | // Default behaviour for schedule clause. |
| 2962 | CGM.getOpenMPRuntime().getDefaultScheduleAndChunk( |
| 2963 | *this, S, ScheduleKind.Schedule, ChunkExpr); |
| 2964 | } |
| 2965 | bool HasChunkSizeOne = false; |
| 2966 | llvm::Value *Chunk = nullptr; |
| 2967 | if (ChunkExpr) { |
| 2968 | Chunk = EmitScalarExpr(ChunkExpr); |
| 2969 | Chunk = EmitScalarConversion(Chunk, ChunkExpr->getType(), |
| 2970 | S.getIterationVariable()->getType(), |
| 2971 | S.getBeginLoc()); |
| 2972 | Expr::EvalResult Result; |
| 2973 | if (ChunkExpr->EvaluateAsInt(Result, getContext())) { |
| 2974 | llvm::APSInt EvaluatedChunk = Result.Val.getInt(); |
| 2975 | HasChunkSizeOne = (EvaluatedChunk.getLimitedValue() == 1); |
| 2976 | } |
| 2977 | } |
| 2978 | const unsigned IVSize = getContext().getTypeSize(IVExpr->getType()); |
| 2979 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 2980 | // OpenMP 4.5, 2.7.1 Loop Construct, Description. |
| 2981 | // If the static schedule kind is specified or if the ordered clause is |
| 2982 | // specified, and if no monotonic modifier is specified, the effect will |
| 2983 | // be as if the monotonic modifier was specified. |
| 2984 | bool StaticChunkedOne = RT.isStaticChunked(ScheduleKind.Schedule, |
| 2985 | /* Chunked */ Chunk != nullptr) && HasChunkSizeOne && |
| 2986 | isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()); |
| 2987 | bool IsMonotonic = |
| 2988 | Ordered || |
| 2989 | ((ScheduleKind.Schedule == OMPC_SCHEDULE_static || |
| 2990 | ScheduleKind.Schedule == OMPC_SCHEDULE_unknown) && |
| 2991 | !(ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_nonmonotonic || |
| 2992 | ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_nonmonotonic)) || |
| 2993 | ScheduleKind.M1 == OMPC_SCHEDULE_MODIFIER_monotonic || |
| 2994 | ScheduleKind.M2 == OMPC_SCHEDULE_MODIFIER_monotonic; |
| 2995 | if ((RT.isStaticNonchunked(ScheduleKind.Schedule, |
| 2996 | /* Chunked */ Chunk != nullptr) || |
| 2997 | StaticChunkedOne) && |
| 2998 | !Ordered) { |
| 2999 | JumpDest LoopExit = |
| 3000 | getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit" )); |
| 3001 | emitCommonSimdLoop( |
| 3002 | *this, S, |
| 3003 | [&S, IsMonotonic](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3004 | if (isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 3005 | CGF.EmitOMPSimdInit(S, IsMonotonic); |
| 3006 | } else if (const auto *C = S.getSingleClause<OMPOrderClause>()) { |
| 3007 | if (C->getKind() == OMPC_ORDER_concurrent) |
| 3008 | CGF.LoopStack.setParallel(/*Enable=*/true); |
| 3009 | } |
| 3010 | }, |
| 3011 | [IVSize, IVSigned, Ordered, IL, LB, UB, ST, StaticChunkedOne, Chunk, |
| 3012 | &S, ScheduleKind, LoopExit, |
| 3013 | &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3014 | // OpenMP [2.7.1, Loop Construct, Description, table 2-1] |
| 3015 | // When no chunk_size is specified, the iteration space is divided |
| 3016 | // into chunks that are approximately equal in size, and at most |
| 3017 | // one chunk is distributed to each thread. Note that the size of |
| 3018 | // the chunks is unspecified in this case. |
| 3019 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 3020 | IVSize, IVSigned, Ordered, IL.getAddress(CGF), |
| 3021 | LB.getAddress(CGF), UB.getAddress(CGF), ST.getAddress(CGF), |
| 3022 | StaticChunkedOne ? Chunk : nullptr); |
| 3023 | CGF.CGM.getOpenMPRuntime().emitForStaticInit( |
| 3024 | CGF, S.getBeginLoc(), S.getDirectiveKind(), ScheduleKind, |
| 3025 | StaticInit); |
| 3026 | // UB = min(UB, GlobalUB); |
| 3027 | if (!StaticChunkedOne) |
| 3028 | CGF.EmitIgnoredExpr(S.getEnsureUpperBound()); |
| 3029 | // IV = LB; |
| 3030 | CGF.EmitIgnoredExpr(S.getInit()); |
| 3031 | // For unchunked static schedule generate: |
| 3032 | // |
| 3033 | // while (idx <= UB) { |
| 3034 | // BODY; |
| 3035 | // ++idx; |
| 3036 | // } |
| 3037 | // |
| 3038 | // For static schedule with chunk one: |
| 3039 | // |
| 3040 | // while (IV <= PrevUB) { |
| 3041 | // BODY; |
| 3042 | // IV += ST; |
| 3043 | // } |
| 3044 | CGF.EmitOMPInnerLoop( |
| 3045 | S, LoopScope.requiresCleanups(), |
| 3046 | StaticChunkedOne ? S.getCombinedParForInDistCond() |
| 3047 | : S.getCond(), |
| 3048 | StaticChunkedOne ? S.getDistInc() : S.getInc(), |
| 3049 | [&S, LoopExit](CodeGenFunction &CGF) { |
| 3050 | emitOMPLoopBodyWithStopPoint(CGF, S, LoopExit); |
| 3051 | }, |
| 3052 | [](CodeGenFunction &) {}); |
| 3053 | }); |
| 3054 | EmitBlock(LoopExit.getBlock()); |
| 3055 | // Tell the runtime we are done. |
| 3056 | auto &&CodeGen = [&S](CodeGenFunction &CGF) { |
| 3057 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(), |
| 3058 | S.getDirectiveKind()); |
| 3059 | }; |
| 3060 | OMPCancelStack.emitExit(*this, S.getDirectiveKind(), CodeGen); |
| 3061 | } else { |
| 3062 | // Emit the outer loop, which requests its work chunk [LB..UB] from |
| 3063 | // runtime and runs the inner loop to process it. |
| 3064 | const OMPLoopArguments LoopArguments( |
| 3065 | LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this), |
| 3066 | IL.getAddress(*this), Chunk, EUB); |
| 3067 | EmitOMPForOuterLoop(ScheduleKind, IsMonotonic, S, LoopScope, Ordered, |
| 3068 | LoopArguments, CGDispatchBounds); |
| 3069 | } |
| 3070 | if (isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 3071 | EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) { |
| 3072 | return CGF.Builder.CreateIsNotNull( |
| 3073 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 3074 | }); |
| 3075 | } |
| 3076 | EmitOMPReductionClauseFinal( |
| 3077 | S, /*ReductionKind=*/isOpenMPSimdDirective(S.getDirectiveKind()) |
| 3078 | ? /*Parallel and Simd*/ OMPD_parallel_for_simd |
| 3079 | : /*Parallel only*/ OMPD_parallel); |
| 3080 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 3081 | emitPostUpdateForReductionClause( |
| 3082 | *this, S, [IL, &S](CodeGenFunction &CGF) { |
| 3083 | return CGF.Builder.CreateIsNotNull( |
| 3084 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 3085 | }); |
| 3086 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 3087 | if (HasLastprivateClause) |
| 3088 | EmitOMPLastprivateClauseFinal( |
| 3089 | S, isOpenMPSimdDirective(S.getDirectiveKind()), |
| 3090 | Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc()))); |
| 3091 | } |
| 3092 | EmitOMPLinearClauseFinal(S, [IL, &S](CodeGenFunction &CGF) { |
| 3093 | return CGF.Builder.CreateIsNotNull( |
| 3094 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 3095 | }); |
| 3096 | DoacrossCleanupScope.ForceCleanup(); |
| 3097 | // We're now done with the loop, so jump to the continuation block. |
| 3098 | if (ContBlock) { |
| 3099 | EmitBranch(ContBlock); |
| 3100 | EmitBlock(ContBlock, /*IsFinished=*/true); |
| 3101 | } |
| 3102 | } |
| 3103 | return HasLastprivateClause; |
| 3104 | } |
| 3105 | |
| 3106 | /// The following two functions generate expressions for the loop lower |
| 3107 | /// and upper bounds in case of static and dynamic (dispatch) schedule |
| 3108 | /// of the associated 'for' or 'distribute' loop. |
| 3109 | static std::pair<LValue, LValue> |
| 3110 | emitForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 3111 | const auto &LS = cast<OMPLoopDirective>(S); |
| 3112 | LValue LB = |
| 3113 | EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getLowerBoundVariable())); |
| 3114 | LValue UB = |
| 3115 | EmitOMPHelperVar(CGF, cast<DeclRefExpr>(LS.getUpperBoundVariable())); |
| 3116 | return {LB, UB}; |
| 3117 | } |
| 3118 | |
| 3119 | /// When dealing with dispatch schedules (e.g. dynamic, guided) we do not |
| 3120 | /// consider the lower and upper bound expressions generated by the |
| 3121 | /// worksharing loop support, but we use 0 and the iteration space size as |
| 3122 | /// constants |
| 3123 | static std::pair<llvm::Value *, llvm::Value *> |
| 3124 | emitDispatchForLoopBounds(CodeGenFunction &CGF, const OMPExecutableDirective &S, |
| 3125 | Address LB, Address UB) { |
| 3126 | const auto &LS = cast<OMPLoopDirective>(S); |
| 3127 | const Expr *IVExpr = LS.getIterationVariable(); |
| 3128 | const unsigned IVSize = CGF.getContext().getTypeSize(IVExpr->getType()); |
| 3129 | llvm::Value *LBVal = CGF.Builder.getIntN(IVSize, 0); |
| 3130 | llvm::Value *UBVal = CGF.EmitScalarExpr(LS.getLastIteration()); |
| 3131 | return {LBVal, UBVal}; |
| 3132 | } |
| 3133 | |
| 3134 | /// Emits the code for the directive with inscan reductions. |
| 3135 | /// The code is the following: |
| 3136 | /// \code |
| 3137 | /// size num_iters = <num_iters>; |
| 3138 | /// <type> buffer[num_iters]; |
| 3139 | /// #pragma omp ... |
| 3140 | /// for (i: 0..<num_iters>) { |
| 3141 | /// <input phase>; |
| 3142 | /// buffer[i] = red; |
| 3143 | /// } |
| 3144 | /// for (int k = 0; k != ceil(log2(num_iters)); ++k) |
| 3145 | /// for (size cnt = last_iter; cnt >= pow(2, k); --k) |
| 3146 | /// buffer[i] op= buffer[i-pow(2,k)]; |
| 3147 | /// #pragma omp ... |
| 3148 | /// for (0..<num_iters>) { |
| 3149 | /// red = InclusiveScan ? buffer[i] : buffer[i-1]; |
| 3150 | /// <scan phase>; |
| 3151 | /// } |
| 3152 | /// \endcode |
| 3153 | static void emitScanBasedDirective( |
| 3154 | CodeGenFunction &CGF, const OMPLoopDirective &S, |
| 3155 | llvm::function_ref<llvm::Value *(CodeGenFunction &)> NumIteratorsGen, |
| 3156 | llvm::function_ref<void(CodeGenFunction &)> FirstGen, |
| 3157 | llvm::function_ref<void(CodeGenFunction &)> SecondGen) { |
| 3158 | llvm::Value *OMPScanNumIterations = CGF.Builder.CreateIntCast( |
| 3159 | NumIteratorsGen(CGF), CGF.SizeTy, /*isSigned=*/false); |
| 3160 | SmallVector<const Expr *, 4> Shareds; |
| 3161 | SmallVector<const Expr *, 4> Privates; |
| 3162 | SmallVector<const Expr *, 4> ReductionOps; |
| 3163 | SmallVector<const Expr *, 4> LHSs; |
| 3164 | SmallVector<const Expr *, 4> RHSs; |
| 3165 | SmallVector<const Expr *, 4> CopyOps; |
| 3166 | SmallVector<const Expr *, 4> CopyArrayTemps; |
| 3167 | SmallVector<const Expr *, 4> CopyArrayElems; |
| 3168 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 3169 | assert(C->getModifier() == OMPC_REDUCTION_inscan && |
| 3170 | "Only inscan reductions are expected." ); |
| 3171 | Shareds.append(C->varlist_begin(), C->varlist_end()); |
| 3172 | Privates.append(C->privates().begin(), C->privates().end()); |
| 3173 | ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end()); |
| 3174 | LHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 3175 | RHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 3176 | CopyOps.append(C->copy_ops().begin(), C->copy_ops().end()); |
| 3177 | CopyArrayTemps.append(C->copy_array_temps().begin(), |
| 3178 | C->copy_array_temps().end()); |
| 3179 | CopyArrayElems.append(C->copy_array_elems().begin(), |
| 3180 | C->copy_array_elems().end()); |
| 3181 | } |
| 3182 | { |
| 3183 | // Emit buffers for each reduction variables. |
| 3184 | // ReductionCodeGen is required to emit correctly the code for array |
| 3185 | // reductions. |
| 3186 | ReductionCodeGen RedCG(Shareds, Shareds, Privates, ReductionOps); |
| 3187 | unsigned Count = 0; |
| 3188 | auto *ITA = CopyArrayTemps.begin(); |
| 3189 | for (const Expr *IRef : Privates) { |
| 3190 | const auto *PrivateVD = cast<VarDecl>(cast<DeclRefExpr>(IRef)->getDecl()); |
| 3191 | // Emit variably modified arrays, used for arrays/array sections |
| 3192 | // reductions. |
| 3193 | if (PrivateVD->getType()->isVariablyModifiedType()) { |
| 3194 | RedCG.emitSharedOrigLValue(CGF, Count); |
| 3195 | RedCG.emitAggregateType(CGF, Count); |
| 3196 | } |
| 3197 | CodeGenFunction::OpaqueValueMapping DimMapping( |
| 3198 | CGF, |
| 3199 | cast<OpaqueValueExpr>( |
| 3200 | cast<VariableArrayType>((*ITA)->getType()->getAsArrayTypeUnsafe()) |
| 3201 | ->getSizeExpr()), |
| 3202 | RValue::get(OMPScanNumIterations)); |
| 3203 | // Emit temp buffer. |
| 3204 | CGF.EmitVarDecl(*cast<VarDecl>(cast<DeclRefExpr>(*ITA)->getDecl())); |
| 3205 | ++ITA; |
| 3206 | ++Count; |
| 3207 | } |
| 3208 | } |
| 3209 | CodeGenFunction::ParentLoopDirectiveForScanRegion ScanRegion(CGF, S); |
| 3210 | { |
| 3211 | // Emit loop with input phase: |
| 3212 | // #pragma omp ... |
| 3213 | // for (i: 0..<num_iters>) { |
| 3214 | // <input phase>; |
| 3215 | // buffer[i] = red; |
| 3216 | // } |
| 3217 | CGF.OMPFirstScanLoop = true; |
| 3218 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 3219 | FirstGen(CGF); |
| 3220 | } |
| 3221 | // Emit prefix reduction: |
| 3222 | // for (int k = 0; k <= ceil(log2(n)); ++k) |
| 3223 | llvm::BasicBlock *InputBB = CGF.Builder.GetInsertBlock(); |
| 3224 | llvm::BasicBlock *LoopBB = CGF.createBasicBlock("omp.outer.log.scan.body" ); |
| 3225 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock("omp.outer.log.scan.exit" ); |
| 3226 | llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::log2, CGF.DoubleTy); |
| 3227 | llvm::Value *Arg = |
| 3228 | CGF.Builder.CreateUIToFP(OMPScanNumIterations, CGF.DoubleTy); |
| 3229 | llvm::Value *LogVal = CGF.EmitNounwindRuntimeCall(F, Arg); |
| 3230 | F = CGF.CGM.getIntrinsic(llvm::Intrinsic::ceil, CGF.DoubleTy); |
| 3231 | LogVal = CGF.EmitNounwindRuntimeCall(F, LogVal); |
| 3232 | LogVal = CGF.Builder.CreateFPToUI(LogVal, CGF.IntTy); |
| 3233 | llvm::Value *NMin1 = CGF.Builder.CreateNUWSub( |
| 3234 | OMPScanNumIterations, llvm::ConstantInt::get(CGF.SizeTy, 1)); |
| 3235 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF, S.getBeginLoc()); |
| 3236 | CGF.EmitBlock(LoopBB); |
| 3237 | auto *Counter = CGF.Builder.CreatePHI(CGF.IntTy, 2); |
| 3238 | // size pow2k = 1; |
| 3239 | auto *Pow2K = CGF.Builder.CreatePHI(CGF.SizeTy, 2); |
| 3240 | Counter->addIncoming(llvm::ConstantInt::get(CGF.IntTy, 0), InputBB); |
| 3241 | Pow2K->addIncoming(llvm::ConstantInt::get(CGF.SizeTy, 1), InputBB); |
| 3242 | // for (size i = n - 1; i >= 2 ^ k; --i) |
| 3243 | // tmp[i] op= tmp[i-pow2k]; |
| 3244 | llvm::BasicBlock *InnerLoopBB = |
| 3245 | CGF.createBasicBlock("omp.inner.log.scan.body" ); |
| 3246 | llvm::BasicBlock *InnerExitBB = |
| 3247 | CGF.createBasicBlock("omp.inner.log.scan.exit" ); |
| 3248 | llvm::Value *CmpI = CGF.Builder.CreateICmpUGE(NMin1, Pow2K); |
| 3249 | CGF.Builder.CreateCondBr(CmpI, InnerLoopBB, InnerExitBB); |
| 3250 | CGF.EmitBlock(InnerLoopBB); |
| 3251 | auto *IVal = CGF.Builder.CreatePHI(CGF.SizeTy, 2); |
| 3252 | IVal->addIncoming(NMin1, LoopBB); |
| 3253 | { |
| 3254 | CodeGenFunction::OMPPrivateScope PrivScope(CGF); |
| 3255 | auto *ILHS = LHSs.begin(); |
| 3256 | auto *IRHS = RHSs.begin(); |
| 3257 | for (const Expr *CopyArrayElem : CopyArrayElems) { |
| 3258 | const auto *LHSVD = cast<VarDecl>(cast<DeclRefExpr>(*ILHS)->getDecl()); |
| 3259 | const auto *RHSVD = cast<VarDecl>(cast<DeclRefExpr>(*IRHS)->getDecl()); |
| 3260 | Address LHSAddr = Address::invalid(); |
| 3261 | { |
| 3262 | CodeGenFunction::OpaqueValueMapping IdxMapping( |
| 3263 | CGF, |
| 3264 | cast<OpaqueValueExpr>( |
| 3265 | cast<ArraySubscriptExpr>(CopyArrayElem)->getIdx()), |
| 3266 | RValue::get(IVal)); |
| 3267 | LHSAddr = CGF.EmitLValue(CopyArrayElem).getAddress(CGF); |
| 3268 | } |
| 3269 | PrivScope.addPrivate(LHSVD, [LHSAddr]() { return LHSAddr; }); |
| 3270 | Address RHSAddr = Address::invalid(); |
| 3271 | { |
| 3272 | llvm::Value *OffsetIVal = CGF.Builder.CreateNUWSub(IVal, Pow2K); |
| 3273 | CodeGenFunction::OpaqueValueMapping IdxMapping( |
| 3274 | CGF, |
| 3275 | cast<OpaqueValueExpr>( |
| 3276 | cast<ArraySubscriptExpr>(CopyArrayElem)->getIdx()), |
| 3277 | RValue::get(OffsetIVal)); |
| 3278 | RHSAddr = CGF.EmitLValue(CopyArrayElem).getAddress(CGF); |
| 3279 | } |
| 3280 | PrivScope.addPrivate(RHSVD, [RHSAddr]() { return RHSAddr; }); |
| 3281 | ++ILHS; |
| 3282 | ++IRHS; |
| 3283 | } |
| 3284 | PrivScope.Privatize(); |
| 3285 | CGF.CGM.getOpenMPRuntime().emitReduction( |
| 3286 | CGF, S.getEndLoc(), Privates, LHSs, RHSs, ReductionOps, |
| 3287 | {/*WithNowait=*/true, /*SimpleReduction=*/true, OMPD_unknown}); |
| 3288 | } |
| 3289 | llvm::Value *NextIVal = |
| 3290 | CGF.Builder.CreateNUWSub(IVal, llvm::ConstantInt::get(CGF.SizeTy, 1)); |
| 3291 | IVal->addIncoming(NextIVal, CGF.Builder.GetInsertBlock()); |
| 3292 | CmpI = CGF.Builder.CreateICmpUGE(NextIVal, Pow2K); |
| 3293 | CGF.Builder.CreateCondBr(CmpI, InnerLoopBB, InnerExitBB); |
| 3294 | CGF.EmitBlock(InnerExitBB); |
| 3295 | llvm::Value *Next = |
| 3296 | CGF.Builder.CreateNUWAdd(Counter, llvm::ConstantInt::get(CGF.IntTy, 1)); |
| 3297 | Counter->addIncoming(Next, CGF.Builder.GetInsertBlock()); |
| 3298 | // pow2k <<= 1; |
| 3299 | llvm::Value *NextPow2K = CGF.Builder.CreateShl(Pow2K, 1, "" , /*HasNUW=*/true); |
| 3300 | Pow2K->addIncoming(NextPow2K, CGF.Builder.GetInsertBlock()); |
| 3301 | llvm::Value *Cmp = CGF.Builder.CreateICmpNE(Next, LogVal); |
| 3302 | CGF.Builder.CreateCondBr(Cmp, LoopBB, ExitBB); |
| 3303 | auto DL1 = ApplyDebugLocation::CreateDefaultArtificial(CGF, S.getEndLoc()); |
| 3304 | CGF.EmitBlock(ExitBB); |
| 3305 | |
| 3306 | CGF.OMPFirstScanLoop = false; |
| 3307 | SecondGen(CGF); |
| 3308 | } |
| 3309 | |
| 3310 | static bool emitWorksharingDirective(CodeGenFunction &CGF, |
| 3311 | const OMPLoopDirective &S, |
| 3312 | bool HasCancel) { |
| 3313 | bool HasLastprivates; |
| 3314 | if (llvm::any_of(S.getClausesOfKind<OMPReductionClause>(), |
| 3315 | [](const OMPReductionClause *C) { |
| 3316 | return C->getModifier() == OMPC_REDUCTION_inscan; |
| 3317 | })) { |
| 3318 | const auto &&NumIteratorsGen = [&S](CodeGenFunction &CGF) { |
| 3319 | CodeGenFunction::OMPLocalDeclMapRAII Scope(CGF); |
| 3320 | OMPLoopScope LoopScope(CGF, S); |
| 3321 | return CGF.EmitScalarExpr(S.getNumIterations()); |
| 3322 | }; |
| 3323 | const auto &&FirstGen = [&S, HasCancel](CodeGenFunction &CGF) { |
| 3324 | CodeGenFunction::OMPCancelStackRAII CancelRegion( |
| 3325 | CGF, S.getDirectiveKind(), HasCancel); |
| 3326 | (void)CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), |
| 3327 | emitForLoopBounds, |
| 3328 | emitDispatchForLoopBounds); |
| 3329 | // Emit an implicit barrier at the end. |
| 3330 | CGF.CGM.getOpenMPRuntime().emitBarrierCall(CGF, S.getBeginLoc(), |
| 3331 | OMPD_for); |
| 3332 | }; |
| 3333 | const auto &&SecondGen = [&S, HasCancel, |
| 3334 | &HasLastprivates](CodeGenFunction &CGF) { |
| 3335 | CodeGenFunction::OMPCancelStackRAII CancelRegion( |
| 3336 | CGF, S.getDirectiveKind(), HasCancel); |
| 3337 | HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), |
| 3338 | emitForLoopBounds, |
| 3339 | emitDispatchForLoopBounds); |
| 3340 | }; |
| 3341 | emitScanBasedDirective(CGF, S, NumIteratorsGen, FirstGen, SecondGen); |
| 3342 | } else { |
| 3343 | CodeGenFunction::OMPCancelStackRAII CancelRegion(CGF, S.getDirectiveKind(), |
| 3344 | HasCancel); |
| 3345 | HasLastprivates = CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), |
| 3346 | emitForLoopBounds, |
| 3347 | emitDispatchForLoopBounds); |
| 3348 | } |
| 3349 | return HasLastprivates; |
| 3350 | } |
| 3351 | |
| 3352 | void CodeGenFunction::EmitOMPForDirective(const OMPForDirective &S) { |
| 3353 | bool HasLastprivates = false; |
| 3354 | auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF, |
| 3355 | PrePostActionTy &) { |
| 3356 | HasLastprivates = emitWorksharingDirective(CGF, S, S.hasCancel()); |
| 3357 | }; |
| 3358 | { |
| 3359 | auto LPCRegion = |
| 3360 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3361 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 3362 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_for, CodeGen, |
| 3363 | S.hasCancel()); |
| 3364 | } |
| 3365 | |
| 3366 | // Emit an implicit barrier at the end. |
| 3367 | if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates) |
| 3368 | CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for); |
| 3369 | // Check for outer lastprivate conditional update. |
| 3370 | checkForLastprivateConditionalUpdate(*this, S); |
| 3371 | } |
| 3372 | |
| 3373 | void CodeGenFunction::EmitOMPForSimdDirective(const OMPForSimdDirective &S) { |
| 3374 | bool HasLastprivates = false; |
| 3375 | auto &&CodeGen = [&S, &HasLastprivates](CodeGenFunction &CGF, |
| 3376 | PrePostActionTy &) { |
| 3377 | HasLastprivates = emitWorksharingDirective(CGF, S, /*HasCancel=*/false); |
| 3378 | }; |
| 3379 | { |
| 3380 | auto LPCRegion = |
| 3381 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3382 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 3383 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_simd, CodeGen); |
| 3384 | } |
| 3385 | |
| 3386 | // Emit an implicit barrier at the end. |
| 3387 | if (!S.getSingleClause<OMPNowaitClause>() || HasLastprivates) |
| 3388 | CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_for); |
| 3389 | // Check for outer lastprivate conditional update. |
| 3390 | checkForLastprivateConditionalUpdate(*this, S); |
| 3391 | } |
| 3392 | |
| 3393 | static LValue createSectionLVal(CodeGenFunction &CGF, QualType Ty, |
| 3394 | const Twine &Name, |
| 3395 | llvm::Value *Init = nullptr) { |
| 3396 | LValue LVal = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty, Name), Ty); |
| 3397 | if (Init) |
| 3398 | CGF.EmitStoreThroughLValue(RValue::get(Init), LVal, /*isInit*/ true); |
| 3399 | return LVal; |
| 3400 | } |
| 3401 | |
| 3402 | void CodeGenFunction::EmitSections(const OMPExecutableDirective &S) { |
| 3403 | const Stmt *CapturedStmt = S.getInnermostCapturedStmt()->getCapturedStmt(); |
| 3404 | const auto *CS = dyn_cast<CompoundStmt>(CapturedStmt); |
| 3405 | bool HasLastprivates = false; |
| 3406 | auto &&CodeGen = [&S, CapturedStmt, CS, |
| 3407 | &HasLastprivates](CodeGenFunction &CGF, PrePostActionTy &) { |
| 3408 | const ASTContext &C = CGF.getContext(); |
| 3409 | QualType KmpInt32Ty = |
| 3410 | C.getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1); |
| 3411 | // Emit helper vars inits. |
| 3412 | LValue LB = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.lb." , |
| 3413 | CGF.Builder.getInt32(0)); |
| 3414 | llvm::ConstantInt *GlobalUBVal = CS != nullptr |
| 3415 | ? CGF.Builder.getInt32(CS->size() - 1) |
| 3416 | : CGF.Builder.getInt32(0); |
| 3417 | LValue UB = |
| 3418 | createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.ub." , GlobalUBVal); |
| 3419 | LValue ST = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.st." , |
| 3420 | CGF.Builder.getInt32(1)); |
| 3421 | LValue IL = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.il." , |
| 3422 | CGF.Builder.getInt32(0)); |
| 3423 | // Loop counter. |
| 3424 | LValue IV = createSectionLVal(CGF, KmpInt32Ty, ".omp.sections.iv." ); |
| 3425 | OpaqueValueExpr IVRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue); |
| 3426 | CodeGenFunction::OpaqueValueMapping OpaqueIV(CGF, &IVRefExpr, IV); |
| 3427 | OpaqueValueExpr UBRefExpr(S.getBeginLoc(), KmpInt32Ty, VK_LValue); |
| 3428 | CodeGenFunction::OpaqueValueMapping OpaqueUB(CGF, &UBRefExpr, UB); |
| 3429 | // Generate condition for loop. |
| 3430 | BinaryOperator *Cond = BinaryOperator::Create( |
| 3431 | C, &IVRefExpr, &UBRefExpr, BO_LE, C.BoolTy, VK_RValue, OK_Ordinary, |
| 3432 | S.getBeginLoc(), FPOptionsOverride()); |
| 3433 | // Increment for loop counter. |
| 3434 | UnaryOperator *Inc = UnaryOperator::Create( |
| 3435 | C, &IVRefExpr, UO_PreInc, KmpInt32Ty, VK_RValue, OK_Ordinary, |
| 3436 | S.getBeginLoc(), true, FPOptionsOverride()); |
| 3437 | auto &&BodyGen = [CapturedStmt, CS, &S, &IV](CodeGenFunction &CGF) { |
| 3438 | // Iterate through all sections and emit a switch construct: |
| 3439 | // switch (IV) { |
| 3440 | // case 0: |
| 3441 | // <SectionStmt[0]>; |
| 3442 | // break; |
| 3443 | // ... |
| 3444 | // case <NumSection> - 1: |
| 3445 | // <SectionStmt[<NumSection> - 1]>; |
| 3446 | // break; |
| 3447 | // } |
| 3448 | // .omp.sections.exit: |
| 3449 | llvm::BasicBlock *ExitBB = CGF.createBasicBlock(".omp.sections.exit" ); |
| 3450 | llvm::SwitchInst *SwitchStmt = |
| 3451 | CGF.Builder.CreateSwitch(CGF.EmitLoadOfScalar(IV, S.getBeginLoc()), |
| 3452 | ExitBB, CS == nullptr ? 1 : CS->size()); |
| 3453 | if (CS) { |
| 3454 | unsigned CaseNumber = 0; |
| 3455 | for (const Stmt *SubStmt : CS->children()) { |
| 3456 | auto CaseBB = CGF.createBasicBlock(".omp.sections.case" ); |
| 3457 | CGF.EmitBlock(CaseBB); |
| 3458 | SwitchStmt->addCase(CGF.Builder.getInt32(CaseNumber), CaseBB); |
| 3459 | CGF.EmitStmt(SubStmt); |
| 3460 | CGF.EmitBranch(ExitBB); |
| 3461 | ++CaseNumber; |
| 3462 | } |
| 3463 | } else { |
| 3464 | llvm::BasicBlock *CaseBB = CGF.createBasicBlock(".omp.sections.case" ); |
| 3465 | CGF.EmitBlock(CaseBB); |
| 3466 | SwitchStmt->addCase(CGF.Builder.getInt32(0), CaseBB); |
| 3467 | CGF.EmitStmt(CapturedStmt); |
| 3468 | CGF.EmitBranch(ExitBB); |
| 3469 | } |
| 3470 | CGF.EmitBlock(ExitBB, /*IsFinished=*/true); |
| 3471 | }; |
| 3472 | |
| 3473 | CodeGenFunction::OMPPrivateScope LoopScope(CGF); |
| 3474 | if (CGF.EmitOMPFirstprivateClause(S, LoopScope)) { |
| 3475 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 3476 | // initialization of firstprivate variables and post-update of lastprivate |
| 3477 | // variables. |
| 3478 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 3479 | CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false, |
| 3480 | /*ForceSimpleCall=*/true); |
| 3481 | } |
| 3482 | CGF.EmitOMPPrivateClause(S, LoopScope); |
| 3483 | CGOpenMPRuntime::LastprivateConditionalRAII LPCRegion(CGF, S, IV); |
| 3484 | HasLastprivates = CGF.EmitOMPLastprivateClauseInit(S, LoopScope); |
| 3485 | CGF.EmitOMPReductionClauseInit(S, LoopScope); |
| 3486 | (void)LoopScope.Privatize(); |
| 3487 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 3488 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S); |
| 3489 | |
| 3490 | // Emit static non-chunked loop. |
| 3491 | OpenMPScheduleTy ScheduleKind; |
| 3492 | ScheduleKind.Schedule = OMPC_SCHEDULE_static; |
| 3493 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 3494 | /*IVSize=*/32, /*IVSigned=*/true, /*Ordered=*/false, IL.getAddress(CGF), |
| 3495 | LB.getAddress(CGF), UB.getAddress(CGF), ST.getAddress(CGF)); |
| 3496 | CGF.CGM.getOpenMPRuntime().emitForStaticInit( |
| 3497 | CGF, S.getBeginLoc(), S.getDirectiveKind(), ScheduleKind, StaticInit); |
| 3498 | // UB = min(UB, GlobalUB); |
| 3499 | llvm::Value *UBVal = CGF.EmitLoadOfScalar(UB, S.getBeginLoc()); |
| 3500 | llvm::Value *MinUBGlobalUB = CGF.Builder.CreateSelect( |
| 3501 | CGF.Builder.CreateICmpSLT(UBVal, GlobalUBVal), UBVal, GlobalUBVal); |
| 3502 | CGF.EmitStoreOfScalar(MinUBGlobalUB, UB); |
| 3503 | // IV = LB; |
| 3504 | CGF.EmitStoreOfScalar(CGF.EmitLoadOfScalar(LB, S.getBeginLoc()), IV); |
| 3505 | // while (idx <= UB) { BODY; ++idx; } |
| 3506 | CGF.EmitOMPInnerLoop(S, /*RequiresCleanup=*/false, Cond, Inc, BodyGen, |
| 3507 | [](CodeGenFunction &) {}); |
| 3508 | // Tell the runtime we are done. |
| 3509 | auto &&CodeGen = [&S](CodeGenFunction &CGF) { |
| 3510 | CGF.CGM.getOpenMPRuntime().emitForStaticFinish(CGF, S.getEndLoc(), |
| 3511 | S.getDirectiveKind()); |
| 3512 | }; |
| 3513 | CGF.OMPCancelStack.emitExit(CGF, S.getDirectiveKind(), CodeGen); |
| 3514 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel); |
| 3515 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 3516 | emitPostUpdateForReductionClause(CGF, S, [IL, &S](CodeGenFunction &CGF) { |
| 3517 | return CGF.Builder.CreateIsNotNull( |
| 3518 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 3519 | }); |
| 3520 | |
| 3521 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 3522 | if (HasLastprivates) |
| 3523 | CGF.EmitOMPLastprivateClauseFinal( |
| 3524 | S, /*NoFinals=*/false, |
| 3525 | CGF.Builder.CreateIsNotNull( |
| 3526 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc()))); |
| 3527 | }; |
| 3528 | |
| 3529 | bool HasCancel = false; |
| 3530 | if (auto *OSD = dyn_cast<OMPSectionsDirective>(&S)) |
| 3531 | HasCancel = OSD->hasCancel(); |
| 3532 | else if (auto *OPSD = dyn_cast<OMPParallelSectionsDirective>(&S)) |
| 3533 | HasCancel = OPSD->hasCancel(); |
| 3534 | OMPCancelStackRAII CancelRegion(*this, S.getDirectiveKind(), HasCancel); |
| 3535 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_sections, CodeGen, |
| 3536 | HasCancel); |
| 3537 | // Emit barrier for lastprivates only if 'sections' directive has 'nowait' |
| 3538 | // clause. Otherwise the barrier will be generated by the codegen for the |
| 3539 | // directive. |
| 3540 | if (HasLastprivates && S.getSingleClause<OMPNowaitClause>()) { |
| 3541 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 3542 | // initialization of firstprivate variables. |
| 3543 | CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), |
| 3544 | OMPD_unknown); |
| 3545 | } |
| 3546 | } |
| 3547 | |
| 3548 | void CodeGenFunction::EmitOMPSectionsDirective(const OMPSectionsDirective &S) { |
| 3549 | { |
| 3550 | auto LPCRegion = |
| 3551 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3552 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 3553 | EmitSections(S); |
| 3554 | } |
| 3555 | // Emit an implicit barrier at the end. |
| 3556 | if (!S.getSingleClause<OMPNowaitClause>()) { |
| 3557 | CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), |
| 3558 | OMPD_sections); |
| 3559 | } |
| 3560 | // Check for outer lastprivate conditional update. |
| 3561 | checkForLastprivateConditionalUpdate(*this, S); |
| 3562 | } |
| 3563 | |
| 3564 | void CodeGenFunction::EmitOMPSectionDirective(const OMPSectionDirective &S) { |
| 3565 | LexicalScope Scope(*this, S.getSourceRange()); |
| 3566 | EmitStopPoint(&S); |
| 3567 | EmitStmt(S.getAssociatedStmt()); |
| 3568 | } |
| 3569 | |
| 3570 | void CodeGenFunction::EmitOMPSingleDirective(const OMPSingleDirective &S) { |
| 3571 | llvm::SmallVector<const Expr *, 8> CopyprivateVars; |
| 3572 | llvm::SmallVector<const Expr *, 8> DestExprs; |
| 3573 | llvm::SmallVector<const Expr *, 8> SrcExprs; |
| 3574 | llvm::SmallVector<const Expr *, 8> AssignmentOps; |
| 3575 | // Check if there are any 'copyprivate' clauses associated with this |
| 3576 | // 'single' construct. |
| 3577 | // Build a list of copyprivate variables along with helper expressions |
| 3578 | // (<source>, <destination>, <destination>=<source> expressions) |
| 3579 | for (const auto *C : S.getClausesOfKind<OMPCopyprivateClause>()) { |
| 3580 | CopyprivateVars.append(C->varlists().begin(), C->varlists().end()); |
| 3581 | DestExprs.append(C->destination_exprs().begin(), |
| 3582 | C->destination_exprs().end()); |
| 3583 | SrcExprs.append(C->source_exprs().begin(), C->source_exprs().end()); |
| 3584 | AssignmentOps.append(C->assignment_ops().begin(), |
| 3585 | C->assignment_ops().end()); |
| 3586 | } |
| 3587 | // Emit code for 'single' region along with 'copyprivate' clauses |
| 3588 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3589 | Action.Enter(CGF); |
| 3590 | OMPPrivateScope SingleScope(CGF); |
| 3591 | (void)CGF.EmitOMPFirstprivateClause(S, SingleScope); |
| 3592 | CGF.EmitOMPPrivateClause(S, SingleScope); |
| 3593 | (void)SingleScope.Privatize(); |
| 3594 | CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 3595 | }; |
| 3596 | { |
| 3597 | auto LPCRegion = |
| 3598 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3599 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 3600 | CGM.getOpenMPRuntime().emitSingleRegion(*this, CodeGen, S.getBeginLoc(), |
| 3601 | CopyprivateVars, DestExprs, |
| 3602 | SrcExprs, AssignmentOps); |
| 3603 | } |
| 3604 | // Emit an implicit barrier at the end (to avoid data race on firstprivate |
| 3605 | // init or if no 'nowait' clause was specified and no 'copyprivate' clause). |
| 3606 | if (!S.getSingleClause<OMPNowaitClause>() && CopyprivateVars.empty()) { |
| 3607 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 3608 | *this, S.getBeginLoc(), |
| 3609 | S.getSingleClause<OMPNowaitClause>() ? OMPD_unknown : OMPD_single); |
| 3610 | } |
| 3611 | // Check for outer lastprivate conditional update. |
| 3612 | checkForLastprivateConditionalUpdate(*this, S); |
| 3613 | } |
| 3614 | |
| 3615 | static void emitMaster(CodeGenFunction &CGF, const OMPExecutableDirective &S) { |
| 3616 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3617 | Action.Enter(CGF); |
| 3618 | CGF.EmitStmt(S.getRawStmt()); |
| 3619 | }; |
| 3620 | CGF.CGM.getOpenMPRuntime().emitMasterRegion(CGF, CodeGen, S.getBeginLoc()); |
| 3621 | } |
| 3622 | |
| 3623 | void CodeGenFunction::EmitOMPMasterDirective(const OMPMasterDirective &S) { |
| 3624 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 3625 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 3626 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 3627 | |
| 3628 | const Stmt *MasterRegionBodyStmt = S.getAssociatedStmt(); |
| 3629 | |
| 3630 | auto FiniCB = [this](InsertPointTy IP) { |
| 3631 | OMPBuilderCBHelpers::FinalizeOMPRegion(*this, IP); |
| 3632 | }; |
| 3633 | |
| 3634 | auto BodyGenCB = [MasterRegionBodyStmt, this](InsertPointTy AllocaIP, |
| 3635 | InsertPointTy CodeGenIP, |
| 3636 | llvm::BasicBlock &FiniBB) { |
| 3637 | OMPBuilderCBHelpers::InlinedRegionBodyRAII IRB(*this, AllocaIP, FiniBB); |
| 3638 | OMPBuilderCBHelpers::EmitOMPRegionBody(*this, MasterRegionBodyStmt, |
| 3639 | CodeGenIP, FiniBB); |
| 3640 | }; |
| 3641 | |
| 3642 | LexicalScope Scope(*this, S.getSourceRange()); |
| 3643 | EmitStopPoint(&S); |
| 3644 | Builder.restoreIP(OMPBuilder.createMaster(Builder, BodyGenCB, FiniCB)); |
| 3645 | |
| 3646 | return; |
| 3647 | } |
| 3648 | LexicalScope Scope(*this, S.getSourceRange()); |
| 3649 | EmitStopPoint(&S); |
| 3650 | emitMaster(*this, S); |
| 3651 | } |
| 3652 | |
| 3653 | void CodeGenFunction::EmitOMPCriticalDirective(const OMPCriticalDirective &S) { |
| 3654 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 3655 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 3656 | using InsertPointTy = llvm::OpenMPIRBuilder::InsertPointTy; |
| 3657 | |
| 3658 | const Stmt *CriticalRegionBodyStmt = S.getAssociatedStmt(); |
| 3659 | const Expr *Hint = nullptr; |
| 3660 | if (const auto *HintClause = S.getSingleClause<OMPHintClause>()) |
| 3661 | Hint = HintClause->getHint(); |
| 3662 | |
| 3663 | // TODO: This is slightly different from what's currently being done in |
| 3664 | // clang. Fix the Int32Ty to IntPtrTy (pointer width size) when everything |
| 3665 | // about typing is final. |
| 3666 | llvm::Value *HintInst = nullptr; |
| 3667 | if (Hint) |
| 3668 | HintInst = |
| 3669 | Builder.CreateIntCast(EmitScalarExpr(Hint), CGM.Int32Ty, false); |
| 3670 | |
| 3671 | auto FiniCB = [this](InsertPointTy IP) { |
| 3672 | OMPBuilderCBHelpers::FinalizeOMPRegion(*this, IP); |
| 3673 | }; |
| 3674 | |
| 3675 | auto BodyGenCB = [CriticalRegionBodyStmt, this](InsertPointTy AllocaIP, |
| 3676 | InsertPointTy CodeGenIP, |
| 3677 | llvm::BasicBlock &FiniBB) { |
| 3678 | OMPBuilderCBHelpers::InlinedRegionBodyRAII IRB(*this, AllocaIP, FiniBB); |
| 3679 | OMPBuilderCBHelpers::EmitOMPRegionBody(*this, CriticalRegionBodyStmt, |
| 3680 | CodeGenIP, FiniBB); |
| 3681 | }; |
| 3682 | |
| 3683 | LexicalScope Scope(*this, S.getSourceRange()); |
| 3684 | EmitStopPoint(&S); |
| 3685 | Builder.restoreIP(OMPBuilder.createCritical( |
| 3686 | Builder, BodyGenCB, FiniCB, S.getDirectiveName().getAsString(), |
| 3687 | HintInst)); |
| 3688 | |
| 3689 | return; |
| 3690 | } |
| 3691 | |
| 3692 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3693 | Action.Enter(CGF); |
| 3694 | CGF.EmitStmt(S.getAssociatedStmt()); |
| 3695 | }; |
| 3696 | const Expr *Hint = nullptr; |
| 3697 | if (const auto *HintClause = S.getSingleClause<OMPHintClause>()) |
| 3698 | Hint = HintClause->getHint(); |
| 3699 | LexicalScope Scope(*this, S.getSourceRange()); |
| 3700 | EmitStopPoint(&S); |
| 3701 | CGM.getOpenMPRuntime().emitCriticalRegion(*this, |
| 3702 | S.getDirectiveName().getAsString(), |
| 3703 | CodeGen, S.getBeginLoc(), Hint); |
| 3704 | } |
| 3705 | |
| 3706 | void CodeGenFunction::EmitOMPParallelForDirective( |
| 3707 | const OMPParallelForDirective &S) { |
| 3708 | // Emit directive as a combined directive that consists of two implicit |
| 3709 | // directives: 'parallel' with 'for' directive. |
| 3710 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3711 | Action.Enter(CGF); |
| 3712 | (void)emitWorksharingDirective(CGF, S, S.hasCancel()); |
| 3713 | }; |
| 3714 | { |
| 3715 | auto LPCRegion = |
| 3716 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3717 | emitCommonOMPParallelDirective(*this, S, OMPD_for, CodeGen, |
| 3718 | emitEmptyBoundParameters); |
| 3719 | } |
| 3720 | // Check for outer lastprivate conditional update. |
| 3721 | checkForLastprivateConditionalUpdate(*this, S); |
| 3722 | } |
| 3723 | |
| 3724 | void CodeGenFunction::EmitOMPParallelForSimdDirective( |
| 3725 | const OMPParallelForSimdDirective &S) { |
| 3726 | // Emit directive as a combined directive that consists of two implicit |
| 3727 | // directives: 'parallel' with 'for' directive. |
| 3728 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3729 | Action.Enter(CGF); |
| 3730 | (void)emitWorksharingDirective(CGF, S, /*HasCancel=*/false); |
| 3731 | }; |
| 3732 | { |
| 3733 | auto LPCRegion = |
| 3734 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3735 | emitCommonOMPParallelDirective(*this, S, OMPD_for_simd, CodeGen, |
| 3736 | emitEmptyBoundParameters); |
| 3737 | } |
| 3738 | // Check for outer lastprivate conditional update. |
| 3739 | checkForLastprivateConditionalUpdate(*this, S); |
| 3740 | } |
| 3741 | |
| 3742 | void CodeGenFunction::EmitOMPParallelMasterDirective( |
| 3743 | const OMPParallelMasterDirective &S) { |
| 3744 | // Emit directive as a combined directive that consists of two implicit |
| 3745 | // directives: 'parallel' with 'master' directive. |
| 3746 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3747 | Action.Enter(CGF); |
| 3748 | OMPPrivateScope PrivateScope(CGF); |
| 3749 | bool Copyins = CGF.EmitOMPCopyinClause(S); |
| 3750 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 3751 | if (Copyins) { |
| 3752 | // Emit implicit barrier to synchronize threads and avoid data races on |
| 3753 | // propagation master's thread values of threadprivate variables to local |
| 3754 | // instances of that variables of all other implicit threads. |
| 3755 | CGF.CGM.getOpenMPRuntime().emitBarrierCall( |
| 3756 | CGF, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false, |
| 3757 | /*ForceSimpleCall=*/true); |
| 3758 | } |
| 3759 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 3760 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 3761 | (void)PrivateScope.Privatize(); |
| 3762 | emitMaster(CGF, S); |
| 3763 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel); |
| 3764 | }; |
| 3765 | { |
| 3766 | auto LPCRegion = |
| 3767 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3768 | emitCommonOMPParallelDirective(*this, S, OMPD_master, CodeGen, |
| 3769 | emitEmptyBoundParameters); |
| 3770 | emitPostUpdateForReductionClause(*this, S, |
| 3771 | [](CodeGenFunction &) { return nullptr; }); |
| 3772 | } |
| 3773 | // Check for outer lastprivate conditional update. |
| 3774 | checkForLastprivateConditionalUpdate(*this, S); |
| 3775 | } |
| 3776 | |
| 3777 | void CodeGenFunction::EmitOMPParallelSectionsDirective( |
| 3778 | const OMPParallelSectionsDirective &S) { |
| 3779 | // Emit directive as a combined directive that consists of two implicit |
| 3780 | // directives: 'parallel' with 'sections' directive. |
| 3781 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 3782 | Action.Enter(CGF); |
| 3783 | CGF.EmitSections(S); |
| 3784 | }; |
| 3785 | { |
| 3786 | auto LPCRegion = |
| 3787 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 3788 | emitCommonOMPParallelDirective(*this, S, OMPD_sections, CodeGen, |
| 3789 | emitEmptyBoundParameters); |
| 3790 | } |
| 3791 | // Check for outer lastprivate conditional update. |
| 3792 | checkForLastprivateConditionalUpdate(*this, S); |
| 3793 | } |
| 3794 | |
| 3795 | namespace { |
| 3796 | /// Get the list of variables declared in the context of the untied tasks. |
| 3797 | class CheckVarsEscapingUntiedTaskDeclContext final |
| 3798 | : public ConstStmtVisitor<CheckVarsEscapingUntiedTaskDeclContext> { |
| 3799 | llvm::SmallVector<const VarDecl *, 4> PrivateDecls; |
| 3800 | |
| 3801 | public: |
| 3802 | explicit CheckVarsEscapingUntiedTaskDeclContext() = default; |
| 3803 | virtual ~CheckVarsEscapingUntiedTaskDeclContext() = default; |
| 3804 | void VisitDeclStmt(const DeclStmt *S) { |
| 3805 | if (!S) |
| 3806 | return; |
| 3807 | // Need to privatize only local vars, static locals can be processed as is. |
| 3808 | for (const Decl *D : S->decls()) { |
| 3809 | if (const auto *VD = dyn_cast_or_null<VarDecl>(D)) |
| 3810 | if (VD->hasLocalStorage()) |
| 3811 | PrivateDecls.push_back(VD); |
| 3812 | } |
| 3813 | } |
| 3814 | void VisitOMPExecutableDirective(const OMPExecutableDirective *) { return; } |
| 3815 | void VisitCapturedStmt(const CapturedStmt *) { return; } |
| 3816 | void VisitLambdaExpr(const LambdaExpr *) { return; } |
| 3817 | void VisitBlockExpr(const BlockExpr *) { return; } |
| 3818 | void VisitStmt(const Stmt *S) { |
| 3819 | if (!S) |
| 3820 | return; |
| 3821 | for (const Stmt *Child : S->children()) |
| 3822 | if (Child) |
| 3823 | Visit(Child); |
| 3824 | } |
| 3825 | |
| 3826 | /// Swaps list of vars with the provided one. |
| 3827 | ArrayRef<const VarDecl *> getPrivateDecls() const { return PrivateDecls; } |
| 3828 | }; |
| 3829 | } // anonymous namespace |
| 3830 | |
| 3831 | void CodeGenFunction::EmitOMPTaskBasedDirective( |
| 3832 | const OMPExecutableDirective &S, const OpenMPDirectiveKind CapturedRegion, |
| 3833 | const RegionCodeGenTy &BodyGen, const TaskGenTy &TaskGen, |
| 3834 | OMPTaskDataTy &Data) { |
| 3835 | // Emit outlined function for task construct. |
| 3836 | const CapturedStmt *CS = S.getCapturedStmt(CapturedRegion); |
| 3837 | auto I = CS->getCapturedDecl()->param_begin(); |
| 3838 | auto PartId = std::next(I); |
| 3839 | auto TaskT = std::next(I, 4); |
| 3840 | // Check if the task is final |
| 3841 | if (const auto *Clause = S.getSingleClause<OMPFinalClause>()) { |
| 3842 | // If the condition constant folds and can be elided, try to avoid emitting |
| 3843 | // the condition and the dead arm of the if/else. |
| 3844 | const Expr *Cond = Clause->getCondition(); |
| 3845 | bool CondConstant; |
| 3846 | if (ConstantFoldsToSimpleInteger(Cond, CondConstant)) |
| 3847 | Data.Final.setInt(CondConstant); |
| 3848 | else |
| 3849 | Data.Final.setPointer(EvaluateExprAsBool(Cond)); |
| 3850 | } else { |
| 3851 | // By default the task is not final. |
| 3852 | Data.Final.setInt(/*IntVal=*/false); |
| 3853 | } |
| 3854 | // Check if the task has 'priority' clause. |
| 3855 | if (const auto *Clause = S.getSingleClause<OMPPriorityClause>()) { |
| 3856 | const Expr *Prio = Clause->getPriority(); |
| 3857 | Data.Priority.setInt(/*IntVal=*/true); |
| 3858 | Data.Priority.setPointer(EmitScalarConversion( |
| 3859 | EmitScalarExpr(Prio), Prio->getType(), |
| 3860 | getContext().getIntTypeForBitwidth(/*DestWidth=*/32, /*Signed=*/1), |
| 3861 | Prio->getExprLoc())); |
| 3862 | } |
| 3863 | // The first function argument for tasks is a thread id, the second one is a |
| 3864 | // part id (0 for tied tasks, >=0 for untied task). |
| 3865 | llvm::DenseSet<const VarDecl *> EmittedAsPrivate; |
| 3866 | // Get list of private variables. |
| 3867 | for (const auto *C : S.getClausesOfKind<OMPPrivateClause>()) { |
| 3868 | auto IRef = C->varlist_begin(); |
| 3869 | for (const Expr *IInit : C->private_copies()) { |
| 3870 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 3871 | if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 3872 | Data.PrivateVars.push_back(*IRef); |
| 3873 | Data.PrivateCopies.push_back(IInit); |
| 3874 | } |
| 3875 | ++IRef; |
| 3876 | } |
| 3877 | } |
| 3878 | EmittedAsPrivate.clear(); |
| 3879 | // Get list of firstprivate variables. |
| 3880 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 3881 | auto IRef = C->varlist_begin(); |
| 3882 | auto IElemInitRef = C->inits().begin(); |
| 3883 | for (const Expr *IInit : C->private_copies()) { |
| 3884 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 3885 | if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 3886 | Data.FirstprivateVars.push_back(*IRef); |
| 3887 | Data.FirstprivateCopies.push_back(IInit); |
| 3888 | Data.FirstprivateInits.push_back(*IElemInitRef); |
| 3889 | } |
| 3890 | ++IRef; |
| 3891 | ++IElemInitRef; |
| 3892 | } |
| 3893 | } |
| 3894 | // Get list of lastprivate variables (for taskloops). |
| 3895 | llvm::DenseMap<const VarDecl *, const DeclRefExpr *> LastprivateDstsOrigs; |
| 3896 | for (const auto *C : S.getClausesOfKind<OMPLastprivateClause>()) { |
| 3897 | auto IRef = C->varlist_begin(); |
| 3898 | auto ID = C->destination_exprs().begin(); |
| 3899 | for (const Expr *IInit : C->private_copies()) { |
| 3900 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl()); |
| 3901 | if (EmittedAsPrivate.insert(OrigVD->getCanonicalDecl()).second) { |
| 3902 | Data.LastprivateVars.push_back(*IRef); |
| 3903 | Data.LastprivateCopies.push_back(IInit); |
| 3904 | } |
| 3905 | LastprivateDstsOrigs.insert( |
| 3906 | {cast<VarDecl>(cast<DeclRefExpr>(*ID)->getDecl()), |
| 3907 | cast<DeclRefExpr>(*IRef)}); |
| 3908 | ++IRef; |
| 3909 | ++ID; |
| 3910 | } |
| 3911 | } |
| 3912 | SmallVector<const Expr *, 4> LHSs; |
| 3913 | SmallVector<const Expr *, 4> RHSs; |
| 3914 | for (const auto *C : S.getClausesOfKind<OMPReductionClause>()) { |
| 3915 | Data.ReductionVars.append(C->varlist_begin(), C->varlist_end()); |
| 3916 | Data.ReductionOrigs.append(C->varlist_begin(), C->varlist_end()); |
| 3917 | Data.ReductionCopies.append(C->privates().begin(), C->privates().end()); |
| 3918 | Data.ReductionOps.append(C->reduction_ops().begin(), |
| 3919 | C->reduction_ops().end()); |
| 3920 | LHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 3921 | RHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 3922 | } |
| 3923 | Data.Reductions = CGM.getOpenMPRuntime().emitTaskReductionInit( |
| 3924 | *this, S.getBeginLoc(), LHSs, RHSs, Data); |
| 3925 | // Build list of dependences. |
| 3926 | for (const auto *C : S.getClausesOfKind<OMPDependClause>()) { |
| 3927 | OMPTaskDataTy::DependData &DD = |
| 3928 | Data.Dependences.emplace_back(C->getDependencyKind(), C->getModifier()); |
| 3929 | DD.DepExprs.append(C->varlist_begin(), C->varlist_end()); |
| 3930 | } |
| 3931 | // Get list of local vars for untied tasks. |
| 3932 | if (!Data.Tied) { |
| 3933 | CheckVarsEscapingUntiedTaskDeclContext Checker; |
| 3934 | Checker.Visit(S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 3935 | Data.PrivateLocals.append(Checker.getPrivateDecls().begin(), |
| 3936 | Checker.getPrivateDecls().end()); |
| 3937 | } |
| 3938 | auto &&CodeGen = [&Data, &S, CS, &BodyGen, &LastprivateDstsOrigs, |
| 3939 | CapturedRegion](CodeGenFunction &CGF, |
| 3940 | PrePostActionTy &Action) { |
| 3941 | llvm::DenseMap<CanonicalDeclPtr<const VarDecl>, std::pair<Address, Address>> |
| 3942 | UntiedLocalVars; |
| 3943 | // Set proper addresses for generated private copies. |
| 3944 | OMPPrivateScope Scope(CGF); |
| 3945 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> FirstprivatePtrs; |
| 3946 | if (!Data.PrivateVars.empty() || !Data.FirstprivateVars.empty() || |
| 3947 | !Data.LastprivateVars.empty() || !Data.PrivateLocals.empty()) { |
| 3948 | llvm::FunctionType *CopyFnTy = llvm::FunctionType::get( |
| 3949 | CGF.Builder.getVoidTy(), {CGF.Builder.getInt8PtrTy()}, true); |
| 3950 | enum { PrivatesParam = 2, CopyFnParam = 3 }; |
| 3951 | llvm::Value *CopyFn = CGF.Builder.CreateLoad( |
| 3952 | CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam))); |
| 3953 | llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar( |
| 3954 | CS->getCapturedDecl()->getParam(PrivatesParam))); |
| 3955 | // Map privates. |
| 3956 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs; |
| 3957 | llvm::SmallVector<llvm::Value *, 16> CallArgs; |
| 3958 | CallArgs.push_back(PrivatesPtr); |
| 3959 | for (const Expr *E : Data.PrivateVars) { |
| 3960 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 3961 | Address PrivatePtr = CGF.CreateMemTemp( |
| 3962 | CGF.getContext().getPointerType(E->getType()), ".priv.ptr.addr" ); |
| 3963 | PrivatePtrs.emplace_back(VD, PrivatePtr); |
| 3964 | CallArgs.push_back(PrivatePtr.getPointer()); |
| 3965 | } |
| 3966 | for (const Expr *E : Data.FirstprivateVars) { |
| 3967 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 3968 | Address PrivatePtr = |
| 3969 | CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()), |
| 3970 | ".firstpriv.ptr.addr" ); |
| 3971 | PrivatePtrs.emplace_back(VD, PrivatePtr); |
| 3972 | FirstprivatePtrs.emplace_back(VD, PrivatePtr); |
| 3973 | CallArgs.push_back(PrivatePtr.getPointer()); |
| 3974 | } |
| 3975 | for (const Expr *E : Data.LastprivateVars) { |
| 3976 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 3977 | Address PrivatePtr = |
| 3978 | CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()), |
| 3979 | ".lastpriv.ptr.addr" ); |
| 3980 | PrivatePtrs.emplace_back(VD, PrivatePtr); |
| 3981 | CallArgs.push_back(PrivatePtr.getPointer()); |
| 3982 | } |
| 3983 | for (const VarDecl *VD : Data.PrivateLocals) { |
| 3984 | QualType Ty = VD->getType().getNonReferenceType(); |
| 3985 | if (VD->getType()->isLValueReferenceType()) |
| 3986 | Ty = CGF.getContext().getPointerType(Ty); |
| 3987 | if (isAllocatableDecl(VD)) |
| 3988 | Ty = CGF.getContext().getPointerType(Ty); |
| 3989 | Address PrivatePtr = CGF.CreateMemTemp( |
| 3990 | CGF.getContext().getPointerType(Ty), ".local.ptr.addr" ); |
| 3991 | UntiedLocalVars.try_emplace(VD, PrivatePtr, Address::invalid()); |
| 3992 | CallArgs.push_back(PrivatePtr.getPointer()); |
| 3993 | } |
| 3994 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 3995 | CGF, S.getBeginLoc(), {CopyFnTy, CopyFn}, CallArgs); |
| 3996 | for (const auto &Pair : LastprivateDstsOrigs) { |
| 3997 | const auto *OrigVD = cast<VarDecl>(Pair.second->getDecl()); |
| 3998 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(OrigVD), |
| 3999 | /*RefersToEnclosingVariableOrCapture=*/ |
| 4000 | CGF.CapturedStmtInfo->lookup(OrigVD) != nullptr, |
| 4001 | Pair.second->getType(), VK_LValue, |
| 4002 | Pair.second->getExprLoc()); |
| 4003 | Scope.addPrivate(Pair.first, [&CGF, &DRE]() { |
| 4004 | return CGF.EmitLValue(&DRE).getAddress(CGF); |
| 4005 | }); |
| 4006 | } |
| 4007 | for (const auto &Pair : PrivatePtrs) { |
| 4008 | Address Replacement(CGF.Builder.CreateLoad(Pair.second), |
| 4009 | CGF.getContext().getDeclAlign(Pair.first)); |
| 4010 | Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; }); |
| 4011 | } |
| 4012 | // Adjust mapping for internal locals by mapping actual memory instead of |
| 4013 | // a pointer to this memory. |
| 4014 | for (auto &Pair : UntiedLocalVars) { |
| 4015 | if (isAllocatableDecl(Pair.first)) { |
| 4016 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Pair.second.first); |
| 4017 | Address Replacement(Ptr, CGF.getPointerAlign()); |
| 4018 | Pair.getSecond().first = Replacement; |
| 4019 | Ptr = CGF.Builder.CreateLoad(Replacement); |
| 4020 | Replacement = Address(Ptr, CGF.getContext().getDeclAlign(Pair.first)); |
| 4021 | Pair.getSecond().second = Replacement; |
| 4022 | } else { |
| 4023 | llvm::Value *Ptr = CGF.Builder.CreateLoad(Pair.second.first); |
| 4024 | Address Replacement(Ptr, CGF.getContext().getDeclAlign(Pair.first)); |
| 4025 | Pair.getSecond().first = Replacement; |
| 4026 | } |
| 4027 | } |
| 4028 | } |
| 4029 | if (Data.Reductions) { |
| 4030 | OMPPrivateScope FirstprivateScope(CGF); |
| 4031 | for (const auto &Pair : FirstprivatePtrs) { |
| 4032 | Address Replacement(CGF.Builder.CreateLoad(Pair.second), |
| 4033 | CGF.getContext().getDeclAlign(Pair.first)); |
| 4034 | FirstprivateScope.addPrivate(Pair.first, |
| 4035 | [Replacement]() { return Replacement; }); |
| 4036 | } |
| 4037 | (void)FirstprivateScope.Privatize(); |
| 4038 | OMPLexicalScope LexScope(CGF, S, CapturedRegion); |
| 4039 | ReductionCodeGen RedCG(Data.ReductionVars, Data.ReductionVars, |
| 4040 | Data.ReductionCopies, Data.ReductionOps); |
| 4041 | llvm::Value *ReductionsPtr = CGF.Builder.CreateLoad( |
| 4042 | CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(9))); |
| 4043 | for (unsigned Cnt = 0, E = Data.ReductionVars.size(); Cnt < E; ++Cnt) { |
| 4044 | RedCG.emitSharedOrigLValue(CGF, Cnt); |
| 4045 | RedCG.emitAggregateType(CGF, Cnt); |
| 4046 | // FIXME: This must removed once the runtime library is fixed. |
| 4047 | // Emit required threadprivate variables for |
| 4048 | // initializer/combiner/finalizer. |
| 4049 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(), |
| 4050 | RedCG, Cnt); |
| 4051 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 4052 | CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt)); |
| 4053 | Replacement = |
| 4054 | Address(CGF.EmitScalarConversion( |
| 4055 | Replacement.getPointer(), CGF.getContext().VoidPtrTy, |
| 4056 | CGF.getContext().getPointerType( |
| 4057 | Data.ReductionCopies[Cnt]->getType()), |
| 4058 | Data.ReductionCopies[Cnt]->getExprLoc()), |
| 4059 | Replacement.getAlignment()); |
| 4060 | Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement); |
| 4061 | Scope.addPrivate(RedCG.getBaseDecl(Cnt), |
| 4062 | [Replacement]() { return Replacement; }); |
| 4063 | } |
| 4064 | } |
| 4065 | // Privatize all private variables except for in_reduction items. |
| 4066 | (void)Scope.Privatize(); |
| 4067 | SmallVector<const Expr *, 4> InRedVars; |
| 4068 | SmallVector<const Expr *, 4> InRedPrivs; |
| 4069 | SmallVector<const Expr *, 4> InRedOps; |
| 4070 | SmallVector<const Expr *, 4> TaskgroupDescriptors; |
| 4071 | for (const auto *C : S.getClausesOfKind<OMPInReductionClause>()) { |
| 4072 | auto IPriv = C->privates().begin(); |
| 4073 | auto IRed = C->reduction_ops().begin(); |
| 4074 | auto ITD = C->taskgroup_descriptors().begin(); |
| 4075 | for (const Expr *Ref : C->varlists()) { |
| 4076 | InRedVars.emplace_back(Ref); |
| 4077 | InRedPrivs.emplace_back(*IPriv); |
| 4078 | InRedOps.emplace_back(*IRed); |
| 4079 | TaskgroupDescriptors.emplace_back(*ITD); |
| 4080 | std::advance(IPriv, 1); |
| 4081 | std::advance(IRed, 1); |
| 4082 | std::advance(ITD, 1); |
| 4083 | } |
| 4084 | } |
| 4085 | // Privatize in_reduction items here, because taskgroup descriptors must be |
| 4086 | // privatized earlier. |
| 4087 | OMPPrivateScope InRedScope(CGF); |
| 4088 | if (!InRedVars.empty()) { |
| 4089 | ReductionCodeGen RedCG(InRedVars, InRedVars, InRedPrivs, InRedOps); |
| 4090 | for (unsigned Cnt = 0, E = InRedVars.size(); Cnt < E; ++Cnt) { |
| 4091 | RedCG.emitSharedOrigLValue(CGF, Cnt); |
| 4092 | RedCG.emitAggregateType(CGF, Cnt); |
| 4093 | // The taskgroup descriptor variable is always implicit firstprivate and |
| 4094 | // privatized already during processing of the firstprivates. |
| 4095 | // FIXME: This must removed once the runtime library is fixed. |
| 4096 | // Emit required threadprivate variables for |
| 4097 | // initializer/combiner/finalizer. |
| 4098 | CGF.CGM.getOpenMPRuntime().emitTaskReductionFixups(CGF, S.getBeginLoc(), |
| 4099 | RedCG, Cnt); |
| 4100 | llvm::Value *ReductionsPtr; |
| 4101 | if (const Expr *TRExpr = TaskgroupDescriptors[Cnt]) { |
| 4102 | ReductionsPtr = CGF.EmitLoadOfScalar(CGF.EmitLValue(TRExpr), |
| 4103 | TRExpr->getExprLoc()); |
| 4104 | } else { |
| 4105 | ReductionsPtr = llvm::ConstantPointerNull::get(CGF.VoidPtrTy); |
| 4106 | } |
| 4107 | Address Replacement = CGF.CGM.getOpenMPRuntime().getTaskReductionItem( |
| 4108 | CGF, S.getBeginLoc(), ReductionsPtr, RedCG.getSharedLValue(Cnt)); |
| 4109 | Replacement = Address( |
| 4110 | CGF.EmitScalarConversion( |
| 4111 | Replacement.getPointer(), CGF.getContext().VoidPtrTy, |
| 4112 | CGF.getContext().getPointerType(InRedPrivs[Cnt]->getType()), |
| 4113 | InRedPrivs[Cnt]->getExprLoc()), |
| 4114 | Replacement.getAlignment()); |
| 4115 | Replacement = RedCG.adjustPrivateAddress(CGF, Cnt, Replacement); |
| 4116 | InRedScope.addPrivate(RedCG.getBaseDecl(Cnt), |
| 4117 | [Replacement]() { return Replacement; }); |
| 4118 | } |
| 4119 | } |
| 4120 | (void)InRedScope.Privatize(); |
| 4121 | |
| 4122 | CGOpenMPRuntime::UntiedTaskLocalDeclsRAII (CGF, |
| 4123 | UntiedLocalVars); |
| 4124 | Action.Enter(CGF); |
| 4125 | BodyGen(CGF); |
| 4126 | }; |
| 4127 | llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction( |
| 4128 | S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, Data.Tied, |
| 4129 | Data.NumberOfParts); |
| 4130 | OMPLexicalScope Scope(*this, S, llvm::None, |
| 4131 | !isOpenMPParallelDirective(S.getDirectiveKind()) && |
| 4132 | !isOpenMPSimdDirective(S.getDirectiveKind())); |
| 4133 | TaskGen(*this, OutlinedFn, Data); |
| 4134 | } |
| 4135 | |
| 4136 | static ImplicitParamDecl * |
| 4137 | createImplicitFirstprivateForType(ASTContext &C, OMPTaskDataTy &Data, |
| 4138 | QualType Ty, CapturedDecl *CD, |
| 4139 | SourceLocation Loc) { |
| 4140 | auto *OrigVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty, |
| 4141 | ImplicitParamDecl::Other); |
| 4142 | auto *OrigRef = DeclRefExpr::Create( |
| 4143 | C, NestedNameSpecifierLoc(), SourceLocation(), OrigVD, |
| 4144 | /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue); |
| 4145 | auto *PrivateVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, Ty, |
| 4146 | ImplicitParamDecl::Other); |
| 4147 | auto *PrivateRef = DeclRefExpr::Create( |
| 4148 | C, NestedNameSpecifierLoc(), SourceLocation(), PrivateVD, |
| 4149 | /*RefersToEnclosingVariableOrCapture=*/false, Loc, Ty, VK_LValue); |
| 4150 | QualType ElemType = C.getBaseElementType(Ty); |
| 4151 | auto *InitVD = ImplicitParamDecl::Create(C, CD, Loc, /*Id=*/nullptr, ElemType, |
| 4152 | ImplicitParamDecl::Other); |
| 4153 | auto *InitRef = DeclRefExpr::Create( |
| 4154 | C, NestedNameSpecifierLoc(), SourceLocation(), InitVD, |
| 4155 | /*RefersToEnclosingVariableOrCapture=*/false, Loc, ElemType, VK_LValue); |
| 4156 | PrivateVD->setInitStyle(VarDecl::CInit); |
| 4157 | PrivateVD->setInit(ImplicitCastExpr::Create(C, ElemType, CK_LValueToRValue, |
| 4158 | InitRef, /*BasePath=*/nullptr, |
| 4159 | VK_RValue, FPOptionsOverride())); |
| 4160 | Data.FirstprivateVars.emplace_back(OrigRef); |
| 4161 | Data.FirstprivateCopies.emplace_back(PrivateRef); |
| 4162 | Data.FirstprivateInits.emplace_back(InitRef); |
| 4163 | return OrigVD; |
| 4164 | } |
| 4165 | |
| 4166 | void CodeGenFunction::EmitOMPTargetTaskBasedDirective( |
| 4167 | const OMPExecutableDirective &S, const RegionCodeGenTy &BodyGen, |
| 4168 | OMPTargetDataInfo &InputInfo) { |
| 4169 | // Emit outlined function for task construct. |
| 4170 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_task); |
| 4171 | Address CapturedStruct = GenerateCapturedStmtArgument(*CS); |
| 4172 | QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl()); |
| 4173 | auto I = CS->getCapturedDecl()->param_begin(); |
| 4174 | auto PartId = std::next(I); |
| 4175 | auto TaskT = std::next(I, 4); |
| 4176 | OMPTaskDataTy Data; |
| 4177 | // The task is not final. |
| 4178 | Data.Final.setInt(/*IntVal=*/false); |
| 4179 | // Get list of firstprivate variables. |
| 4180 | for (const auto *C : S.getClausesOfKind<OMPFirstprivateClause>()) { |
| 4181 | auto IRef = C->varlist_begin(); |
| 4182 | auto IElemInitRef = C->inits().begin(); |
| 4183 | for (auto *IInit : C->private_copies()) { |
| 4184 | Data.FirstprivateVars.push_back(*IRef); |
| 4185 | Data.FirstprivateCopies.push_back(IInit); |
| 4186 | Data.FirstprivateInits.push_back(*IElemInitRef); |
| 4187 | ++IRef; |
| 4188 | ++IElemInitRef; |
| 4189 | } |
| 4190 | } |
| 4191 | OMPPrivateScope TargetScope(*this); |
| 4192 | VarDecl *BPVD = nullptr; |
| 4193 | VarDecl *PVD = nullptr; |
| 4194 | VarDecl *SVD = nullptr; |
| 4195 | VarDecl *MVD = nullptr; |
| 4196 | if (InputInfo.NumberOfTargetItems > 0) { |
| 4197 | auto *CD = CapturedDecl::Create( |
| 4198 | getContext(), getContext().getTranslationUnitDecl(), /*NumParams=*/0); |
| 4199 | llvm::APInt ArrSize(/*numBits=*/32, InputInfo.NumberOfTargetItems); |
| 4200 | QualType BaseAndPointerAndMapperType = getContext().getConstantArrayType( |
| 4201 | getContext().VoidPtrTy, ArrSize, nullptr, ArrayType::Normal, |
| 4202 | /*IndexTypeQuals=*/0); |
| 4203 | BPVD = createImplicitFirstprivateForType( |
| 4204 | getContext(), Data, BaseAndPointerAndMapperType, CD, S.getBeginLoc()); |
| 4205 | PVD = createImplicitFirstprivateForType( |
| 4206 | getContext(), Data, BaseAndPointerAndMapperType, CD, S.getBeginLoc()); |
| 4207 | QualType SizesType = getContext().getConstantArrayType( |
| 4208 | getContext().getIntTypeForBitwidth(/*DestWidth=*/64, /*Signed=*/1), |
| 4209 | ArrSize, nullptr, ArrayType::Normal, |
| 4210 | /*IndexTypeQuals=*/0); |
| 4211 | SVD = createImplicitFirstprivateForType(getContext(), Data, SizesType, CD, |
| 4212 | S.getBeginLoc()); |
| 4213 | TargetScope.addPrivate( |
| 4214 | BPVD, [&InputInfo]() { return InputInfo.BasePointersArray; }); |
| 4215 | TargetScope.addPrivate(PVD, |
| 4216 | [&InputInfo]() { return InputInfo.PointersArray; }); |
| 4217 | TargetScope.addPrivate(SVD, |
| 4218 | [&InputInfo]() { return InputInfo.SizesArray; }); |
| 4219 | // If there is no user-defined mapper, the mapper array will be nullptr. In |
| 4220 | // this case, we don't need to privatize it. |
| 4221 | if (!dyn_cast_or_null<llvm::ConstantPointerNull>( |
| 4222 | InputInfo.MappersArray.getPointer())) { |
| 4223 | MVD = createImplicitFirstprivateForType( |
| 4224 | getContext(), Data, BaseAndPointerAndMapperType, CD, S.getBeginLoc()); |
| 4225 | TargetScope.addPrivate(MVD, |
| 4226 | [&InputInfo]() { return InputInfo.MappersArray; }); |
| 4227 | } |
| 4228 | } |
| 4229 | (void)TargetScope.Privatize(); |
| 4230 | // Build list of dependences. |
| 4231 | for (const auto *C : S.getClausesOfKind<OMPDependClause>()) { |
| 4232 | OMPTaskDataTy::DependData &DD = |
| 4233 | Data.Dependences.emplace_back(C->getDependencyKind(), C->getModifier()); |
| 4234 | DD.DepExprs.append(C->varlist_begin(), C->varlist_end()); |
| 4235 | } |
| 4236 | auto &&CodeGen = [&Data, &S, CS, &BodyGen, BPVD, PVD, SVD, MVD, |
| 4237 | &InputInfo](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4238 | // Set proper addresses for generated private copies. |
| 4239 | OMPPrivateScope Scope(CGF); |
| 4240 | if (!Data.FirstprivateVars.empty()) { |
| 4241 | llvm::FunctionType *CopyFnTy = llvm::FunctionType::get( |
| 4242 | CGF.Builder.getVoidTy(), {CGF.Builder.getInt8PtrTy()}, true); |
| 4243 | enum { PrivatesParam = 2, CopyFnParam = 3 }; |
| 4244 | llvm::Value *CopyFn = CGF.Builder.CreateLoad( |
| 4245 | CGF.GetAddrOfLocalVar(CS->getCapturedDecl()->getParam(CopyFnParam))); |
| 4246 | llvm::Value *PrivatesPtr = CGF.Builder.CreateLoad(CGF.GetAddrOfLocalVar( |
| 4247 | CS->getCapturedDecl()->getParam(PrivatesParam))); |
| 4248 | // Map privates. |
| 4249 | llvm::SmallVector<std::pair<const VarDecl *, Address>, 16> PrivatePtrs; |
| 4250 | llvm::SmallVector<llvm::Value *, 16> CallArgs; |
| 4251 | CallArgs.push_back(PrivatesPtr); |
| 4252 | for (const Expr *E : Data.FirstprivateVars) { |
| 4253 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 4254 | Address PrivatePtr = |
| 4255 | CGF.CreateMemTemp(CGF.getContext().getPointerType(E->getType()), |
| 4256 | ".firstpriv.ptr.addr" ); |
| 4257 | PrivatePtrs.emplace_back(VD, PrivatePtr); |
| 4258 | CallArgs.push_back(PrivatePtr.getPointer()); |
| 4259 | } |
| 4260 | CGF.CGM.getOpenMPRuntime().emitOutlinedFunctionCall( |
| 4261 | CGF, S.getBeginLoc(), {CopyFnTy, CopyFn}, CallArgs); |
| 4262 | for (const auto &Pair : PrivatePtrs) { |
| 4263 | Address Replacement(CGF.Builder.CreateLoad(Pair.second), |
| 4264 | CGF.getContext().getDeclAlign(Pair.first)); |
| 4265 | Scope.addPrivate(Pair.first, [Replacement]() { return Replacement; }); |
| 4266 | } |
| 4267 | } |
| 4268 | // Privatize all private variables except for in_reduction items. |
| 4269 | (void)Scope.Privatize(); |
| 4270 | if (InputInfo.NumberOfTargetItems > 0) { |
| 4271 | InputInfo.BasePointersArray = CGF.Builder.CreateConstArrayGEP( |
| 4272 | CGF.GetAddrOfLocalVar(BPVD), /*Index=*/0); |
| 4273 | InputInfo.PointersArray = CGF.Builder.CreateConstArrayGEP( |
| 4274 | CGF.GetAddrOfLocalVar(PVD), /*Index=*/0); |
| 4275 | InputInfo.SizesArray = CGF.Builder.CreateConstArrayGEP( |
| 4276 | CGF.GetAddrOfLocalVar(SVD), /*Index=*/0); |
| 4277 | // If MVD is nullptr, the mapper array is not privatized |
| 4278 | if (MVD) |
| 4279 | InputInfo.MappersArray = CGF.Builder.CreateConstArrayGEP( |
| 4280 | CGF.GetAddrOfLocalVar(MVD), /*Index=*/0); |
| 4281 | } |
| 4282 | |
| 4283 | Action.Enter(CGF); |
| 4284 | OMPLexicalScope LexScope(CGF, S, OMPD_task, /*EmitPreInitStmt=*/false); |
| 4285 | BodyGen(CGF); |
| 4286 | }; |
| 4287 | llvm::Function *OutlinedFn = CGM.getOpenMPRuntime().emitTaskOutlinedFunction( |
| 4288 | S, *I, *PartId, *TaskT, S.getDirectiveKind(), CodeGen, /*Tied=*/true, |
| 4289 | Data.NumberOfParts); |
| 4290 | llvm::APInt TrueOrFalse(32, S.hasClausesOfKind<OMPNowaitClause>() ? 1 : 0); |
| 4291 | IntegerLiteral IfCond(getContext(), TrueOrFalse, |
| 4292 | getContext().getIntTypeForBitwidth(32, /*Signed=*/0), |
| 4293 | SourceLocation()); |
| 4294 | |
| 4295 | CGM.getOpenMPRuntime().emitTaskCall(*this, S.getBeginLoc(), S, OutlinedFn, |
| 4296 | SharedsTy, CapturedStruct, &IfCond, Data); |
| 4297 | } |
| 4298 | |
| 4299 | void CodeGenFunction::EmitOMPTaskDirective(const OMPTaskDirective &S) { |
| 4300 | // Emit outlined function for task construct. |
| 4301 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_task); |
| 4302 | Address CapturedStruct = GenerateCapturedStmtArgument(*CS); |
| 4303 | QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl()); |
| 4304 | const Expr *IfCond = nullptr; |
| 4305 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 4306 | if (C->getNameModifier() == OMPD_unknown || |
| 4307 | C->getNameModifier() == OMPD_task) { |
| 4308 | IfCond = C->getCondition(); |
| 4309 | break; |
| 4310 | } |
| 4311 | } |
| 4312 | |
| 4313 | OMPTaskDataTy Data; |
| 4314 | // Check if we should emit tied or untied task. |
| 4315 | Data.Tied = !S.getSingleClause<OMPUntiedClause>(); |
| 4316 | auto &&BodyGen = [CS](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4317 | CGF.EmitStmt(CS->getCapturedStmt()); |
| 4318 | }; |
| 4319 | auto &&TaskGen = [&S, SharedsTy, CapturedStruct, |
| 4320 | IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn, |
| 4321 | const OMPTaskDataTy &Data) { |
| 4322 | CGF.CGM.getOpenMPRuntime().emitTaskCall(CGF, S.getBeginLoc(), S, OutlinedFn, |
| 4323 | SharedsTy, CapturedStruct, IfCond, |
| 4324 | Data); |
| 4325 | }; |
| 4326 | auto LPCRegion = |
| 4327 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 4328 | EmitOMPTaskBasedDirective(S, OMPD_task, BodyGen, TaskGen, Data); |
| 4329 | } |
| 4330 | |
| 4331 | void CodeGenFunction::EmitOMPTaskyieldDirective( |
| 4332 | const OMPTaskyieldDirective &S) { |
| 4333 | CGM.getOpenMPRuntime().emitTaskyieldCall(*this, S.getBeginLoc()); |
| 4334 | } |
| 4335 | |
| 4336 | void CodeGenFunction::EmitOMPBarrierDirective(const OMPBarrierDirective &S) { |
| 4337 | CGM.getOpenMPRuntime().emitBarrierCall(*this, S.getBeginLoc(), OMPD_barrier); |
| 4338 | } |
| 4339 | |
| 4340 | void CodeGenFunction::EmitOMPTaskwaitDirective(const OMPTaskwaitDirective &S) { |
| 4341 | CGM.getOpenMPRuntime().emitTaskwaitCall(*this, S.getBeginLoc()); |
| 4342 | } |
| 4343 | |
| 4344 | void CodeGenFunction::EmitOMPTaskgroupDirective( |
| 4345 | const OMPTaskgroupDirective &S) { |
| 4346 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 4347 | Action.Enter(CGF); |
| 4348 | if (const Expr *E = S.getReductionRef()) { |
| 4349 | SmallVector<const Expr *, 4> LHSs; |
| 4350 | SmallVector<const Expr *, 4> RHSs; |
| 4351 | OMPTaskDataTy Data; |
| 4352 | for (const auto *C : S.getClausesOfKind<OMPTaskReductionClause>()) { |
| 4353 | Data.ReductionVars.append(C->varlist_begin(), C->varlist_end()); |
| 4354 | Data.ReductionOrigs.append(C->varlist_begin(), C->varlist_end()); |
| 4355 | Data.ReductionCopies.append(C->privates().begin(), C->privates().end()); |
| 4356 | Data.ReductionOps.append(C->reduction_ops().begin(), |
| 4357 | C->reduction_ops().end()); |
| 4358 | LHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 4359 | RHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 4360 | } |
| 4361 | llvm::Value *ReductionDesc = |
| 4362 | CGF.CGM.getOpenMPRuntime().emitTaskReductionInit(CGF, S.getBeginLoc(), |
| 4363 | LHSs, RHSs, Data); |
| 4364 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 4365 | CGF.EmitVarDecl(*VD); |
| 4366 | CGF.EmitStoreOfScalar(ReductionDesc, CGF.GetAddrOfLocalVar(VD), |
| 4367 | /*Volatile=*/false, E->getType()); |
| 4368 | } |
| 4369 | CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 4370 | }; |
| 4371 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4372 | CGM.getOpenMPRuntime().emitTaskgroupRegion(*this, CodeGen, S.getBeginLoc()); |
| 4373 | } |
| 4374 | |
| 4375 | void CodeGenFunction::EmitOMPFlushDirective(const OMPFlushDirective &S) { |
| 4376 | llvm::AtomicOrdering AO = S.getSingleClause<OMPFlushClause>() |
| 4377 | ? llvm::AtomicOrdering::NotAtomic |
| 4378 | : llvm::AtomicOrdering::AcquireRelease; |
| 4379 | CGM.getOpenMPRuntime().emitFlush( |
| 4380 | *this, |
| 4381 | [&S]() -> ArrayRef<const Expr *> { |
| 4382 | if (const auto *FlushClause = S.getSingleClause<OMPFlushClause>()) |
| 4383 | return llvm::makeArrayRef(FlushClause->varlist_begin(), |
| 4384 | FlushClause->varlist_end()); |
| 4385 | return llvm::None; |
| 4386 | }(), |
| 4387 | S.getBeginLoc(), AO); |
| 4388 | } |
| 4389 | |
| 4390 | void CodeGenFunction::EmitOMPDepobjDirective(const OMPDepobjDirective &S) { |
| 4391 | const auto *DO = S.getSingleClause<OMPDepobjClause>(); |
| 4392 | LValue DOLVal = EmitLValue(DO->getDepobj()); |
| 4393 | if (const auto *DC = S.getSingleClause<OMPDependClause>()) { |
| 4394 | OMPTaskDataTy::DependData Dependencies(DC->getDependencyKind(), |
| 4395 | DC->getModifier()); |
| 4396 | Dependencies.DepExprs.append(DC->varlist_begin(), DC->varlist_end()); |
| 4397 | Address DepAddr = CGM.getOpenMPRuntime().emitDepobjDependClause( |
| 4398 | *this, Dependencies, DC->getBeginLoc()); |
| 4399 | EmitStoreOfScalar(DepAddr.getPointer(), DOLVal); |
| 4400 | return; |
| 4401 | } |
| 4402 | if (const auto *DC = S.getSingleClause<OMPDestroyClause>()) { |
| 4403 | CGM.getOpenMPRuntime().emitDestroyClause(*this, DOLVal, DC->getBeginLoc()); |
| 4404 | return; |
| 4405 | } |
| 4406 | if (const auto *UC = S.getSingleClause<OMPUpdateClause>()) { |
| 4407 | CGM.getOpenMPRuntime().emitUpdateClause( |
| 4408 | *this, DOLVal, UC->getDependencyKind(), UC->getBeginLoc()); |
| 4409 | return; |
| 4410 | } |
| 4411 | } |
| 4412 | |
| 4413 | void CodeGenFunction::EmitOMPScanDirective(const OMPScanDirective &S) { |
| 4414 | if (!OMPParentLoopDirectiveForScan) |
| 4415 | return; |
| 4416 | const OMPExecutableDirective &ParentDir = *OMPParentLoopDirectiveForScan; |
| 4417 | bool IsInclusive = S.hasClausesOfKind<OMPInclusiveClause>(); |
| 4418 | SmallVector<const Expr *, 4> Shareds; |
| 4419 | SmallVector<const Expr *, 4> Privates; |
| 4420 | SmallVector<const Expr *, 4> LHSs; |
| 4421 | SmallVector<const Expr *, 4> RHSs; |
| 4422 | SmallVector<const Expr *, 4> ReductionOps; |
| 4423 | SmallVector<const Expr *, 4> CopyOps; |
| 4424 | SmallVector<const Expr *, 4> CopyArrayTemps; |
| 4425 | SmallVector<const Expr *, 4> CopyArrayElems; |
| 4426 | for (const auto *C : ParentDir.getClausesOfKind<OMPReductionClause>()) { |
| 4427 | if (C->getModifier() != OMPC_REDUCTION_inscan) |
| 4428 | continue; |
| 4429 | Shareds.append(C->varlist_begin(), C->varlist_end()); |
| 4430 | Privates.append(C->privates().begin(), C->privates().end()); |
| 4431 | LHSs.append(C->lhs_exprs().begin(), C->lhs_exprs().end()); |
| 4432 | RHSs.append(C->rhs_exprs().begin(), C->rhs_exprs().end()); |
| 4433 | ReductionOps.append(C->reduction_ops().begin(), C->reduction_ops().end()); |
| 4434 | CopyOps.append(C->copy_ops().begin(), C->copy_ops().end()); |
| 4435 | CopyArrayTemps.append(C->copy_array_temps().begin(), |
| 4436 | C->copy_array_temps().end()); |
| 4437 | CopyArrayElems.append(C->copy_array_elems().begin(), |
| 4438 | C->copy_array_elems().end()); |
| 4439 | } |
| 4440 | if (ParentDir.getDirectiveKind() == OMPD_simd || |
| 4441 | (getLangOpts().OpenMPSimd && |
| 4442 | isOpenMPSimdDirective(ParentDir.getDirectiveKind()))) { |
| 4443 | // For simd directive and simd-based directives in simd only mode, use the |
| 4444 | // following codegen: |
| 4445 | // int x = 0; |
| 4446 | // #pragma omp simd reduction(inscan, +: x) |
| 4447 | // for (..) { |
| 4448 | // <first part> |
| 4449 | // #pragma omp scan inclusive(x) |
| 4450 | // <second part> |
| 4451 | // } |
| 4452 | // is transformed to: |
| 4453 | // int x = 0; |
| 4454 | // for (..) { |
| 4455 | // int x_priv = 0; |
| 4456 | // <first part> |
| 4457 | // x = x_priv + x; |
| 4458 | // x_priv = x; |
| 4459 | // <second part> |
| 4460 | // } |
| 4461 | // and |
| 4462 | // int x = 0; |
| 4463 | // #pragma omp simd reduction(inscan, +: x) |
| 4464 | // for (..) { |
| 4465 | // <first part> |
| 4466 | // #pragma omp scan exclusive(x) |
| 4467 | // <second part> |
| 4468 | // } |
| 4469 | // to |
| 4470 | // int x = 0; |
| 4471 | // for (..) { |
| 4472 | // int x_priv = 0; |
| 4473 | // <second part> |
| 4474 | // int temp = x; |
| 4475 | // x = x_priv + x; |
| 4476 | // x_priv = temp; |
| 4477 | // <first part> |
| 4478 | // } |
| 4479 | llvm::BasicBlock *OMPScanReduce = createBasicBlock("omp.inscan.reduce" ); |
| 4480 | EmitBranch(IsInclusive |
| 4481 | ? OMPScanReduce |
| 4482 | : BreakContinueStack.back().ContinueBlock.getBlock()); |
| 4483 | EmitBlock(OMPScanDispatch); |
| 4484 | { |
| 4485 | // New scope for correct construction/destruction of temp variables for |
| 4486 | // exclusive scan. |
| 4487 | LexicalScope Scope(*this, S.getSourceRange()); |
| 4488 | EmitBranch(IsInclusive ? OMPBeforeScanBlock : OMPAfterScanBlock); |
| 4489 | EmitBlock(OMPScanReduce); |
| 4490 | if (!IsInclusive) { |
| 4491 | // Create temp var and copy LHS value to this temp value. |
| 4492 | // TMP = LHS; |
| 4493 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 4494 | const Expr *PrivateExpr = Privates[I]; |
| 4495 | const Expr *TempExpr = CopyArrayTemps[I]; |
| 4496 | EmitAutoVarDecl( |
| 4497 | *cast<VarDecl>(cast<DeclRefExpr>(TempExpr)->getDecl())); |
| 4498 | LValue DestLVal = EmitLValue(TempExpr); |
| 4499 | LValue SrcLVal = EmitLValue(LHSs[I]); |
| 4500 | EmitOMPCopy(PrivateExpr->getType(), DestLVal.getAddress(*this), |
| 4501 | SrcLVal.getAddress(*this), |
| 4502 | cast<VarDecl>(cast<DeclRefExpr>(LHSs[I])->getDecl()), |
| 4503 | cast<VarDecl>(cast<DeclRefExpr>(RHSs[I])->getDecl()), |
| 4504 | CopyOps[I]); |
| 4505 | } |
| 4506 | } |
| 4507 | CGM.getOpenMPRuntime().emitReduction( |
| 4508 | *this, ParentDir.getEndLoc(), Privates, LHSs, RHSs, ReductionOps, |
| 4509 | {/*WithNowait=*/true, /*SimpleReduction=*/true, OMPD_simd}); |
| 4510 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 4511 | const Expr *PrivateExpr = Privates[I]; |
| 4512 | LValue DestLVal; |
| 4513 | LValue SrcLVal; |
| 4514 | if (IsInclusive) { |
| 4515 | DestLVal = EmitLValue(RHSs[I]); |
| 4516 | SrcLVal = EmitLValue(LHSs[I]); |
| 4517 | } else { |
| 4518 | const Expr *TempExpr = CopyArrayTemps[I]; |
| 4519 | DestLVal = EmitLValue(RHSs[I]); |
| 4520 | SrcLVal = EmitLValue(TempExpr); |
| 4521 | } |
| 4522 | EmitOMPCopy(PrivateExpr->getType(), DestLVal.getAddress(*this), |
| 4523 | SrcLVal.getAddress(*this), |
| 4524 | cast<VarDecl>(cast<DeclRefExpr>(LHSs[I])->getDecl()), |
| 4525 | cast<VarDecl>(cast<DeclRefExpr>(RHSs[I])->getDecl()), |
| 4526 | CopyOps[I]); |
| 4527 | } |
| 4528 | } |
| 4529 | EmitBranch(IsInclusive ? OMPAfterScanBlock : OMPBeforeScanBlock); |
| 4530 | OMPScanExitBlock = IsInclusive |
| 4531 | ? BreakContinueStack.back().ContinueBlock.getBlock() |
| 4532 | : OMPScanReduce; |
| 4533 | EmitBlock(OMPAfterScanBlock); |
| 4534 | return; |
| 4535 | } |
| 4536 | if (!IsInclusive) { |
| 4537 | EmitBranch(BreakContinueStack.back().ContinueBlock.getBlock()); |
| 4538 | EmitBlock(OMPScanExitBlock); |
| 4539 | } |
| 4540 | if (OMPFirstScanLoop) { |
| 4541 | // Emit buffer[i] = red; at the end of the input phase. |
| 4542 | const auto *IVExpr = cast<OMPLoopDirective>(ParentDir) |
| 4543 | .getIterationVariable() |
| 4544 | ->IgnoreParenImpCasts(); |
| 4545 | LValue IdxLVal = EmitLValue(IVExpr); |
| 4546 | llvm::Value *IdxVal = EmitLoadOfScalar(IdxLVal, IVExpr->getExprLoc()); |
| 4547 | IdxVal = Builder.CreateIntCast(IdxVal, SizeTy, /*isSigned=*/false); |
| 4548 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 4549 | const Expr *PrivateExpr = Privates[I]; |
| 4550 | const Expr *OrigExpr = Shareds[I]; |
| 4551 | const Expr *CopyArrayElem = CopyArrayElems[I]; |
| 4552 | OpaqueValueMapping IdxMapping( |
| 4553 | *this, |
| 4554 | cast<OpaqueValueExpr>( |
| 4555 | cast<ArraySubscriptExpr>(CopyArrayElem)->getIdx()), |
| 4556 | RValue::get(IdxVal)); |
| 4557 | LValue DestLVal = EmitLValue(CopyArrayElem); |
| 4558 | LValue SrcLVal = EmitLValue(OrigExpr); |
| 4559 | EmitOMPCopy(PrivateExpr->getType(), DestLVal.getAddress(*this), |
| 4560 | SrcLVal.getAddress(*this), |
| 4561 | cast<VarDecl>(cast<DeclRefExpr>(LHSs[I])->getDecl()), |
| 4562 | cast<VarDecl>(cast<DeclRefExpr>(RHSs[I])->getDecl()), |
| 4563 | CopyOps[I]); |
| 4564 | } |
| 4565 | } |
| 4566 | EmitBranch(BreakContinueStack.back().ContinueBlock.getBlock()); |
| 4567 | if (IsInclusive) { |
| 4568 | EmitBlock(OMPScanExitBlock); |
| 4569 | EmitBranch(BreakContinueStack.back().ContinueBlock.getBlock()); |
| 4570 | } |
| 4571 | EmitBlock(OMPScanDispatch); |
| 4572 | if (!OMPFirstScanLoop) { |
| 4573 | // Emit red = buffer[i]; at the entrance to the scan phase. |
| 4574 | const auto *IVExpr = cast<OMPLoopDirective>(ParentDir) |
| 4575 | .getIterationVariable() |
| 4576 | ->IgnoreParenImpCasts(); |
| 4577 | LValue IdxLVal = EmitLValue(IVExpr); |
| 4578 | llvm::Value *IdxVal = EmitLoadOfScalar(IdxLVal, IVExpr->getExprLoc()); |
| 4579 | IdxVal = Builder.CreateIntCast(IdxVal, SizeTy, /*isSigned=*/false); |
| 4580 | llvm::BasicBlock *ExclusiveExitBB = nullptr; |
| 4581 | if (!IsInclusive) { |
| 4582 | llvm::BasicBlock *ContBB = createBasicBlock("omp.exclusive.dec" ); |
| 4583 | ExclusiveExitBB = createBasicBlock("omp.exclusive.copy.exit" ); |
| 4584 | llvm::Value *Cmp = Builder.CreateIsNull(IdxVal); |
| 4585 | Builder.CreateCondBr(Cmp, ExclusiveExitBB, ContBB); |
| 4586 | EmitBlock(ContBB); |
| 4587 | // Use idx - 1 iteration for exclusive scan. |
| 4588 | IdxVal = Builder.CreateNUWSub(IdxVal, llvm::ConstantInt::get(SizeTy, 1)); |
| 4589 | } |
| 4590 | for (unsigned I = 0, E = CopyArrayElems.size(); I < E; ++I) { |
| 4591 | const Expr *PrivateExpr = Privates[I]; |
| 4592 | const Expr *OrigExpr = Shareds[I]; |
| 4593 | const Expr *CopyArrayElem = CopyArrayElems[I]; |
| 4594 | OpaqueValueMapping IdxMapping( |
| 4595 | *this, |
| 4596 | cast<OpaqueValueExpr>( |
| 4597 | cast<ArraySubscriptExpr>(CopyArrayElem)->getIdx()), |
| 4598 | RValue::get(IdxVal)); |
| 4599 | LValue SrcLVal = EmitLValue(CopyArrayElem); |
| 4600 | LValue DestLVal = EmitLValue(OrigExpr); |
| 4601 | EmitOMPCopy(PrivateExpr->getType(), DestLVal.getAddress(*this), |
| 4602 | SrcLVal.getAddress(*this), |
| 4603 | cast<VarDecl>(cast<DeclRefExpr>(LHSs[I])->getDecl()), |
| 4604 | cast<VarDecl>(cast<DeclRefExpr>(RHSs[I])->getDecl()), |
| 4605 | CopyOps[I]); |
| 4606 | } |
| 4607 | if (!IsInclusive) { |
| 4608 | EmitBlock(ExclusiveExitBB); |
| 4609 | } |
| 4610 | } |
| 4611 | EmitBranch((OMPFirstScanLoop == IsInclusive) ? OMPBeforeScanBlock |
| 4612 | : OMPAfterScanBlock); |
| 4613 | EmitBlock(OMPAfterScanBlock); |
| 4614 | } |
| 4615 | |
| 4616 | void CodeGenFunction::EmitOMPDistributeLoop(const OMPLoopDirective &S, |
| 4617 | const CodeGenLoopTy &CodeGenLoop, |
| 4618 | Expr *IncExpr) { |
| 4619 | // Emit the loop iteration variable. |
| 4620 | const auto *IVExpr = cast<DeclRefExpr>(S.getIterationVariable()); |
| 4621 | const auto *IVDecl = cast<VarDecl>(IVExpr->getDecl()); |
| 4622 | EmitVarDecl(*IVDecl); |
| 4623 | |
| 4624 | // Emit the iterations count variable. |
| 4625 | // If it is not a variable, Sema decided to calculate iterations count on each |
| 4626 | // iteration (e.g., it is foldable into a constant). |
| 4627 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) { |
| 4628 | EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl())); |
| 4629 | // Emit calculation of the iterations count. |
| 4630 | EmitIgnoredExpr(S.getCalcLastIteration()); |
| 4631 | } |
| 4632 | |
| 4633 | CGOpenMPRuntime &RT = CGM.getOpenMPRuntime(); |
| 4634 | |
| 4635 | bool HasLastprivateClause = false; |
| 4636 | // Check pre-condition. |
| 4637 | { |
| 4638 | OMPLoopScope PreInitScope(*this, S); |
| 4639 | // Skip the entire loop if we don't meet the precondition. |
| 4640 | // If the condition constant folds and can be elided, avoid emitting the |
| 4641 | // whole loop. |
| 4642 | bool CondConstant; |
| 4643 | llvm::BasicBlock *ContBlock = nullptr; |
| 4644 | if (ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) { |
| 4645 | if (!CondConstant) |
| 4646 | return; |
| 4647 | } else { |
| 4648 | llvm::BasicBlock *ThenBlock = createBasicBlock("omp.precond.then" ); |
| 4649 | ContBlock = createBasicBlock("omp.precond.end" ); |
| 4650 | emitPreCond(*this, S, S.getPreCond(), ThenBlock, ContBlock, |
| 4651 | getProfileCount(&S)); |
| 4652 | EmitBlock(ThenBlock); |
| 4653 | incrementProfileCounter(&S); |
| 4654 | } |
| 4655 | |
| 4656 | emitAlignedClause(*this, S); |
| 4657 | // Emit 'then' code. |
| 4658 | { |
| 4659 | // Emit helper vars inits. |
| 4660 | |
| 4661 | LValue LB = EmitOMPHelperVar( |
| 4662 | *this, cast<DeclRefExpr>( |
| 4663 | (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 4664 | ? S.getCombinedLowerBoundVariable() |
| 4665 | : S.getLowerBoundVariable()))); |
| 4666 | LValue UB = EmitOMPHelperVar( |
| 4667 | *this, cast<DeclRefExpr>( |
| 4668 | (isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 4669 | ? S.getCombinedUpperBoundVariable() |
| 4670 | : S.getUpperBoundVariable()))); |
| 4671 | LValue ST = |
| 4672 | EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getStrideVariable())); |
| 4673 | LValue IL = |
| 4674 | EmitOMPHelperVar(*this, cast<DeclRefExpr>(S.getIsLastIterVariable())); |
| 4675 | |
| 4676 | OMPPrivateScope LoopScope(*this); |
| 4677 | if (EmitOMPFirstprivateClause(S, LoopScope)) { |
| 4678 | // Emit implicit barrier to synchronize threads and avoid data races |
| 4679 | // on initialization of firstprivate variables and post-update of |
| 4680 | // lastprivate variables. |
| 4681 | CGM.getOpenMPRuntime().emitBarrierCall( |
| 4682 | *this, S.getBeginLoc(), OMPD_unknown, /*EmitChecks=*/false, |
| 4683 | /*ForceSimpleCall=*/true); |
| 4684 | } |
| 4685 | EmitOMPPrivateClause(S, LoopScope); |
| 4686 | if (isOpenMPSimdDirective(S.getDirectiveKind()) && |
| 4687 | !isOpenMPParallelDirective(S.getDirectiveKind()) && |
| 4688 | !isOpenMPTeamsDirective(S.getDirectiveKind())) |
| 4689 | EmitOMPReductionClauseInit(S, LoopScope); |
| 4690 | HasLastprivateClause = EmitOMPLastprivateClauseInit(S, LoopScope); |
| 4691 | EmitOMPPrivateLoopCounters(S, LoopScope); |
| 4692 | (void)LoopScope.Privatize(); |
| 4693 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 4694 | CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(*this, S); |
| 4695 | |
| 4696 | // Detect the distribute schedule kind and chunk. |
| 4697 | llvm::Value *Chunk = nullptr; |
| 4698 | OpenMPDistScheduleClauseKind ScheduleKind = OMPC_DIST_SCHEDULE_unknown; |
| 4699 | if (const auto *C = S.getSingleClause<OMPDistScheduleClause>()) { |
| 4700 | ScheduleKind = C->getDistScheduleKind(); |
| 4701 | if (const Expr *Ch = C->getChunkSize()) { |
| 4702 | Chunk = EmitScalarExpr(Ch); |
| 4703 | Chunk = EmitScalarConversion(Chunk, Ch->getType(), |
| 4704 | S.getIterationVariable()->getType(), |
| 4705 | S.getBeginLoc()); |
| 4706 | } |
| 4707 | } else { |
| 4708 | // Default behaviour for dist_schedule clause. |
| 4709 | CGM.getOpenMPRuntime().getDefaultDistScheduleAndChunk( |
| 4710 | *this, S, ScheduleKind, Chunk); |
| 4711 | } |
| 4712 | const unsigned IVSize = getContext().getTypeSize(IVExpr->getType()); |
| 4713 | const bool IVSigned = IVExpr->getType()->hasSignedIntegerRepresentation(); |
| 4714 | |
| 4715 | // OpenMP [2.10.8, distribute Construct, Description] |
| 4716 | // If dist_schedule is specified, kind must be static. If specified, |
| 4717 | // iterations are divided into chunks of size chunk_size, chunks are |
| 4718 | // assigned to the teams of the league in a round-robin fashion in the |
| 4719 | // order of the team number. When no chunk_size is specified, the |
| 4720 | // iteration space is divided into chunks that are approximately equal |
| 4721 | // in size, and at most one chunk is distributed to each team of the |
| 4722 | // league. The size of the chunks is unspecified in this case. |
| 4723 | bool StaticChunked = RT.isStaticChunked( |
| 4724 | ScheduleKind, /* Chunked */ Chunk != nullptr) && |
| 4725 | isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()); |
| 4726 | if (RT.isStaticNonchunked(ScheduleKind, |
| 4727 | /* Chunked */ Chunk != nullptr) || |
| 4728 | StaticChunked) { |
| 4729 | CGOpenMPRuntime::StaticRTInput StaticInit( |
| 4730 | IVSize, IVSigned, /* Ordered = */ false, IL.getAddress(*this), |
| 4731 | LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this), |
| 4732 | StaticChunked ? Chunk : nullptr); |
| 4733 | RT.emitDistributeStaticInit(*this, S.getBeginLoc(), ScheduleKind, |
| 4734 | StaticInit); |
| 4735 | JumpDest LoopExit = |
| 4736 | getJumpDestInCurrentScope(createBasicBlock("omp.loop.exit" )); |
| 4737 | // UB = min(UB, GlobalUB); |
| 4738 | EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 4739 | ? S.getCombinedEnsureUpperBound() |
| 4740 | : S.getEnsureUpperBound()); |
| 4741 | // IV = LB; |
| 4742 | EmitIgnoredExpr(isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 4743 | ? S.getCombinedInit() |
| 4744 | : S.getInit()); |
| 4745 | |
| 4746 | const Expr *Cond = |
| 4747 | isOpenMPLoopBoundSharingDirective(S.getDirectiveKind()) |
| 4748 | ? S.getCombinedCond() |
| 4749 | : S.getCond(); |
| 4750 | |
| 4751 | if (StaticChunked) |
| 4752 | Cond = S.getCombinedDistCond(); |
| 4753 | |
| 4754 | // For static unchunked schedules generate: |
| 4755 | // |
| 4756 | // 1. For distribute alone, codegen |
| 4757 | // while (idx <= UB) { |
| 4758 | // BODY; |
| 4759 | // ++idx; |
| 4760 | // } |
| 4761 | // |
| 4762 | // 2. When combined with 'for' (e.g. as in 'distribute parallel for') |
| 4763 | // while (idx <= UB) { |
| 4764 | // <CodeGen rest of pragma>(LB, UB); |
| 4765 | // idx += ST; |
| 4766 | // } |
| 4767 | // |
| 4768 | // For static chunk one schedule generate: |
| 4769 | // |
| 4770 | // while (IV <= GlobalUB) { |
| 4771 | // <CodeGen rest of pragma>(LB, UB); |
| 4772 | // LB += ST; |
| 4773 | // UB += ST; |
| 4774 | // UB = min(UB, GlobalUB); |
| 4775 | // IV = LB; |
| 4776 | // } |
| 4777 | // |
| 4778 | emitCommonSimdLoop( |
| 4779 | *this, S, |
| 4780 | [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4781 | if (isOpenMPSimdDirective(S.getDirectiveKind())) |
| 4782 | CGF.EmitOMPSimdInit(S, /*IsMonotonic=*/true); |
| 4783 | }, |
| 4784 | [&S, &LoopScope, Cond, IncExpr, LoopExit, &CodeGenLoop, |
| 4785 | StaticChunked](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4786 | CGF.EmitOMPInnerLoop( |
| 4787 | S, LoopScope.requiresCleanups(), Cond, IncExpr, |
| 4788 | [&S, LoopExit, &CodeGenLoop](CodeGenFunction &CGF) { |
| 4789 | CodeGenLoop(CGF, S, LoopExit); |
| 4790 | }, |
| 4791 | [&S, StaticChunked](CodeGenFunction &CGF) { |
| 4792 | if (StaticChunked) { |
| 4793 | CGF.EmitIgnoredExpr(S.getCombinedNextLowerBound()); |
| 4794 | CGF.EmitIgnoredExpr(S.getCombinedNextUpperBound()); |
| 4795 | CGF.EmitIgnoredExpr(S.getCombinedEnsureUpperBound()); |
| 4796 | CGF.EmitIgnoredExpr(S.getCombinedInit()); |
| 4797 | } |
| 4798 | }); |
| 4799 | }); |
| 4800 | EmitBlock(LoopExit.getBlock()); |
| 4801 | // Tell the runtime we are done. |
| 4802 | RT.emitForStaticFinish(*this, S.getEndLoc(), S.getDirectiveKind()); |
| 4803 | } else { |
| 4804 | // Emit the outer loop, which requests its work chunk [LB..UB] from |
| 4805 | // runtime and runs the inner loop to process it. |
| 4806 | const OMPLoopArguments LoopArguments = { |
| 4807 | LB.getAddress(*this), UB.getAddress(*this), ST.getAddress(*this), |
| 4808 | IL.getAddress(*this), Chunk}; |
| 4809 | EmitOMPDistributeOuterLoop(ScheduleKind, S, LoopScope, LoopArguments, |
| 4810 | CodeGenLoop); |
| 4811 | } |
| 4812 | if (isOpenMPSimdDirective(S.getDirectiveKind())) { |
| 4813 | EmitOMPSimdFinal(S, [IL, &S](CodeGenFunction &CGF) { |
| 4814 | return CGF.Builder.CreateIsNotNull( |
| 4815 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 4816 | }); |
| 4817 | } |
| 4818 | if (isOpenMPSimdDirective(S.getDirectiveKind()) && |
| 4819 | !isOpenMPParallelDirective(S.getDirectiveKind()) && |
| 4820 | !isOpenMPTeamsDirective(S.getDirectiveKind())) { |
| 4821 | EmitOMPReductionClauseFinal(S, OMPD_simd); |
| 4822 | // Emit post-update of the reduction variables if IsLastIter != 0. |
| 4823 | emitPostUpdateForReductionClause( |
| 4824 | *this, S, [IL, &S](CodeGenFunction &CGF) { |
| 4825 | return CGF.Builder.CreateIsNotNull( |
| 4826 | CGF.EmitLoadOfScalar(IL, S.getBeginLoc())); |
| 4827 | }); |
| 4828 | } |
| 4829 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 4830 | if (HasLastprivateClause) { |
| 4831 | EmitOMPLastprivateClauseFinal( |
| 4832 | S, /*NoFinals=*/false, |
| 4833 | Builder.CreateIsNotNull(EmitLoadOfScalar(IL, S.getBeginLoc()))); |
| 4834 | } |
| 4835 | } |
| 4836 | |
| 4837 | // We're now done with the loop, so jump to the continuation block. |
| 4838 | if (ContBlock) { |
| 4839 | EmitBranch(ContBlock); |
| 4840 | EmitBlock(ContBlock, true); |
| 4841 | } |
| 4842 | } |
| 4843 | } |
| 4844 | |
| 4845 | void CodeGenFunction::EmitOMPDistributeDirective( |
| 4846 | const OMPDistributeDirective &S) { |
| 4847 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 4848 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 4849 | }; |
| 4850 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4851 | CGM.getOpenMPRuntime().emitInlinedDirective(*this, OMPD_distribute, CodeGen); |
| 4852 | } |
| 4853 | |
| 4854 | static llvm::Function *emitOutlinedOrderedFunction(CodeGenModule &CGM, |
| 4855 | const CapturedStmt *S, |
| 4856 | SourceLocation Loc) { |
| 4857 | CodeGenFunction CGF(CGM, /*suppressNewContext=*/true); |
| 4858 | CodeGenFunction::CGCapturedStmtInfo CapStmtInfo; |
| 4859 | CGF.CapturedStmtInfo = &CapStmtInfo; |
| 4860 | llvm::Function *Fn = CGF.GenerateOpenMPCapturedStmtFunction(*S, Loc); |
| 4861 | Fn->setDoesNotRecurse(); |
| 4862 | return Fn; |
| 4863 | } |
| 4864 | |
| 4865 | void CodeGenFunction::EmitOMPOrderedDirective(const OMPOrderedDirective &S) { |
| 4866 | if (S.hasClausesOfKind<OMPDependClause>()) { |
| 4867 | assert(!S.hasAssociatedStmt() && |
| 4868 | "No associated statement must be in ordered depend construct." ); |
| 4869 | for (const auto *DC : S.getClausesOfKind<OMPDependClause>()) |
| 4870 | CGM.getOpenMPRuntime().emitDoacrossOrdered(*this, DC); |
| 4871 | return; |
| 4872 | } |
| 4873 | const auto *C = S.getSingleClause<OMPSIMDClause>(); |
| 4874 | auto &&CodeGen = [&S, C, this](CodeGenFunction &CGF, |
| 4875 | PrePostActionTy &Action) { |
| 4876 | const CapturedStmt *CS = S.getInnermostCapturedStmt(); |
| 4877 | if (C) { |
| 4878 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 4879 | CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars); |
| 4880 | llvm::Function *OutlinedFn = |
| 4881 | emitOutlinedOrderedFunction(CGM, CS, S.getBeginLoc()); |
| 4882 | CGM.getOpenMPRuntime().emitOutlinedFunctionCall(CGF, S.getBeginLoc(), |
| 4883 | OutlinedFn, CapturedVars); |
| 4884 | } else { |
| 4885 | Action.Enter(CGF); |
| 4886 | CGF.EmitStmt(CS->getCapturedStmt()); |
| 4887 | } |
| 4888 | }; |
| 4889 | OMPLexicalScope Scope(*this, S, OMPD_unknown); |
| 4890 | CGM.getOpenMPRuntime().emitOrderedRegion(*this, CodeGen, S.getBeginLoc(), !C); |
| 4891 | } |
| 4892 | |
| 4893 | static llvm::Value *convertToScalarValue(CodeGenFunction &CGF, RValue Val, |
| 4894 | QualType SrcType, QualType DestType, |
| 4895 | SourceLocation Loc) { |
| 4896 | assert(CGF.hasScalarEvaluationKind(DestType) && |
| 4897 | "DestType must have scalar evaluation kind." ); |
| 4898 | assert(!Val.isAggregate() && "Must be a scalar or complex." ); |
| 4899 | return Val.isScalar() ? CGF.EmitScalarConversion(Val.getScalarVal(), SrcType, |
| 4900 | DestType, Loc) |
| 4901 | : CGF.EmitComplexToScalarConversion( |
| 4902 | Val.getComplexVal(), SrcType, DestType, Loc); |
| 4903 | } |
| 4904 | |
| 4905 | static CodeGenFunction::ComplexPairTy |
| 4906 | convertToComplexValue(CodeGenFunction &CGF, RValue Val, QualType SrcType, |
| 4907 | QualType DestType, SourceLocation Loc) { |
| 4908 | assert(CGF.getEvaluationKind(DestType) == TEK_Complex && |
| 4909 | "DestType must have complex evaluation kind." ); |
| 4910 | CodeGenFunction::ComplexPairTy ComplexVal; |
| 4911 | if (Val.isScalar()) { |
| 4912 | // Convert the input element to the element type of the complex. |
| 4913 | QualType DestElementType = |
| 4914 | DestType->castAs<ComplexType>()->getElementType(); |
| 4915 | llvm::Value *ScalarVal = CGF.EmitScalarConversion( |
| 4916 | Val.getScalarVal(), SrcType, DestElementType, Loc); |
| 4917 | ComplexVal = CodeGenFunction::ComplexPairTy( |
| 4918 | ScalarVal, llvm::Constant::getNullValue(ScalarVal->getType())); |
| 4919 | } else { |
| 4920 | assert(Val.isComplex() && "Must be a scalar or complex." ); |
| 4921 | QualType SrcElementType = SrcType->castAs<ComplexType>()->getElementType(); |
| 4922 | QualType DestElementType = |
| 4923 | DestType->castAs<ComplexType>()->getElementType(); |
| 4924 | ComplexVal.first = CGF.EmitScalarConversion( |
| 4925 | Val.getComplexVal().first, SrcElementType, DestElementType, Loc); |
| 4926 | ComplexVal.second = CGF.EmitScalarConversion( |
| 4927 | Val.getComplexVal().second, SrcElementType, DestElementType, Loc); |
| 4928 | } |
| 4929 | return ComplexVal; |
| 4930 | } |
| 4931 | |
| 4932 | static void emitSimpleAtomicStore(CodeGenFunction &CGF, llvm::AtomicOrdering AO, |
| 4933 | LValue LVal, RValue RVal) { |
| 4934 | if (LVal.isGlobalReg()) |
| 4935 | CGF.EmitStoreThroughGlobalRegLValue(RVal, LVal); |
| 4936 | else |
| 4937 | CGF.EmitAtomicStore(RVal, LVal, AO, LVal.isVolatile(), /*isInit=*/false); |
| 4938 | } |
| 4939 | |
| 4940 | static RValue emitSimpleAtomicLoad(CodeGenFunction &CGF, |
| 4941 | llvm::AtomicOrdering AO, LValue LVal, |
| 4942 | SourceLocation Loc) { |
| 4943 | if (LVal.isGlobalReg()) |
| 4944 | return CGF.EmitLoadOfLValue(LVal, Loc); |
| 4945 | return CGF.EmitAtomicLoad( |
| 4946 | LVal, Loc, llvm::AtomicCmpXchgInst::getStrongestFailureOrdering(AO), |
| 4947 | LVal.isVolatile()); |
| 4948 | } |
| 4949 | |
| 4950 | void CodeGenFunction::emitOMPSimpleStore(LValue LVal, RValue RVal, |
| 4951 | QualType RValTy, SourceLocation Loc) { |
| 4952 | switch (getEvaluationKind(LVal.getType())) { |
| 4953 | case TEK_Scalar: |
| 4954 | EmitStoreThroughLValue(RValue::get(convertToScalarValue( |
| 4955 | *this, RVal, RValTy, LVal.getType(), Loc)), |
| 4956 | LVal); |
| 4957 | break; |
| 4958 | case TEK_Complex: |
| 4959 | EmitStoreOfComplex( |
| 4960 | convertToComplexValue(*this, RVal, RValTy, LVal.getType(), Loc), LVal, |
| 4961 | /*isInit=*/false); |
| 4962 | break; |
| 4963 | case TEK_Aggregate: |
| 4964 | llvm_unreachable("Must be a scalar or complex." ); |
| 4965 | } |
| 4966 | } |
| 4967 | |
| 4968 | static void emitOMPAtomicReadExpr(CodeGenFunction &CGF, llvm::AtomicOrdering AO, |
| 4969 | const Expr *X, const Expr *V, |
| 4970 | SourceLocation Loc) { |
| 4971 | // v = x; |
| 4972 | assert(V->isLValue() && "V of 'omp atomic read' is not lvalue" ); |
| 4973 | assert(X->isLValue() && "X of 'omp atomic read' is not lvalue" ); |
| 4974 | LValue XLValue = CGF.EmitLValue(X); |
| 4975 | LValue VLValue = CGF.EmitLValue(V); |
| 4976 | RValue Res = emitSimpleAtomicLoad(CGF, AO, XLValue, Loc); |
| 4977 | // OpenMP, 2.17.7, atomic Construct |
| 4978 | // If the read or capture clause is specified and the acquire, acq_rel, or |
| 4979 | // seq_cst clause is specified then the strong flush on exit from the atomic |
| 4980 | // operation is also an acquire flush. |
| 4981 | switch (AO) { |
| 4982 | case llvm::AtomicOrdering::Acquire: |
| 4983 | case llvm::AtomicOrdering::AcquireRelease: |
| 4984 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 4985 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 4986 | llvm::AtomicOrdering::Acquire); |
| 4987 | break; |
| 4988 | case llvm::AtomicOrdering::Monotonic: |
| 4989 | case llvm::AtomicOrdering::Release: |
| 4990 | break; |
| 4991 | case llvm::AtomicOrdering::NotAtomic: |
| 4992 | case llvm::AtomicOrdering::Unordered: |
| 4993 | llvm_unreachable("Unexpected ordering." ); |
| 4994 | } |
| 4995 | CGF.emitOMPSimpleStore(VLValue, Res, X->getType().getNonReferenceType(), Loc); |
| 4996 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, V); |
| 4997 | } |
| 4998 | |
| 4999 | static void emitOMPAtomicWriteExpr(CodeGenFunction &CGF, |
| 5000 | llvm::AtomicOrdering AO, const Expr *X, |
| 5001 | const Expr *E, SourceLocation Loc) { |
| 5002 | // x = expr; |
| 5003 | assert(X->isLValue() && "X of 'omp atomic write' is not lvalue" ); |
| 5004 | emitSimpleAtomicStore(CGF, AO, CGF.EmitLValue(X), CGF.EmitAnyExpr(E)); |
| 5005 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, X); |
| 5006 | // OpenMP, 2.17.7, atomic Construct |
| 5007 | // If the write, update, or capture clause is specified and the release, |
| 5008 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 5009 | // the atomic operation is also a release flush. |
| 5010 | switch (AO) { |
| 5011 | case llvm::AtomicOrdering::Release: |
| 5012 | case llvm::AtomicOrdering::AcquireRelease: |
| 5013 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 5014 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 5015 | llvm::AtomicOrdering::Release); |
| 5016 | break; |
| 5017 | case llvm::AtomicOrdering::Acquire: |
| 5018 | case llvm::AtomicOrdering::Monotonic: |
| 5019 | break; |
| 5020 | case llvm::AtomicOrdering::NotAtomic: |
| 5021 | case llvm::AtomicOrdering::Unordered: |
| 5022 | llvm_unreachable("Unexpected ordering." ); |
| 5023 | } |
| 5024 | } |
| 5025 | |
| 5026 | static std::pair<bool, RValue> emitOMPAtomicRMW(CodeGenFunction &CGF, LValue X, |
| 5027 | RValue Update, |
| 5028 | BinaryOperatorKind BO, |
| 5029 | llvm::AtomicOrdering AO, |
| 5030 | bool IsXLHSInRHSPart) { |
| 5031 | ASTContext &Context = CGF.getContext(); |
| 5032 | // Allow atomicrmw only if 'x' and 'update' are integer values, lvalue for 'x' |
| 5033 | // expression is simple and atomic is allowed for the given type for the |
| 5034 | // target platform. |
| 5035 | if (BO == BO_Comma || !Update.isScalar() || |
| 5036 | !Update.getScalarVal()->getType()->isIntegerTy() || !X.isSimple() || |
| 5037 | (!isa<llvm::ConstantInt>(Update.getScalarVal()) && |
| 5038 | (Update.getScalarVal()->getType() != |
| 5039 | X.getAddress(CGF).getElementType())) || |
| 5040 | !X.getAddress(CGF).getElementType()->isIntegerTy() || |
| 5041 | !Context.getTargetInfo().hasBuiltinAtomic( |
| 5042 | Context.getTypeSize(X.getType()), Context.toBits(X.getAlignment()))) |
| 5043 | return std::make_pair(false, RValue::get(nullptr)); |
| 5044 | |
| 5045 | llvm::AtomicRMWInst::BinOp RMWOp; |
| 5046 | switch (BO) { |
| 5047 | case BO_Add: |
| 5048 | RMWOp = llvm::AtomicRMWInst::Add; |
| 5049 | break; |
| 5050 | case BO_Sub: |
| 5051 | if (!IsXLHSInRHSPart) |
| 5052 | return std::make_pair(false, RValue::get(nullptr)); |
| 5053 | RMWOp = llvm::AtomicRMWInst::Sub; |
| 5054 | break; |
| 5055 | case BO_And: |
| 5056 | RMWOp = llvm::AtomicRMWInst::And; |
| 5057 | break; |
| 5058 | case BO_Or: |
| 5059 | RMWOp = llvm::AtomicRMWInst::Or; |
| 5060 | break; |
| 5061 | case BO_Xor: |
| 5062 | RMWOp = llvm::AtomicRMWInst::Xor; |
| 5063 | break; |
| 5064 | case BO_LT: |
| 5065 | RMWOp = X.getType()->hasSignedIntegerRepresentation() |
| 5066 | ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Min |
| 5067 | : llvm::AtomicRMWInst::Max) |
| 5068 | : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMin |
| 5069 | : llvm::AtomicRMWInst::UMax); |
| 5070 | break; |
| 5071 | case BO_GT: |
| 5072 | RMWOp = X.getType()->hasSignedIntegerRepresentation() |
| 5073 | ? (IsXLHSInRHSPart ? llvm::AtomicRMWInst::Max |
| 5074 | : llvm::AtomicRMWInst::Min) |
| 5075 | : (IsXLHSInRHSPart ? llvm::AtomicRMWInst::UMax |
| 5076 | : llvm::AtomicRMWInst::UMin); |
| 5077 | break; |
| 5078 | case BO_Assign: |
| 5079 | RMWOp = llvm::AtomicRMWInst::Xchg; |
| 5080 | break; |
| 5081 | case BO_Mul: |
| 5082 | case BO_Div: |
| 5083 | case BO_Rem: |
| 5084 | case BO_Shl: |
| 5085 | case BO_Shr: |
| 5086 | case BO_LAnd: |
| 5087 | case BO_LOr: |
| 5088 | return std::make_pair(false, RValue::get(nullptr)); |
| 5089 | case BO_PtrMemD: |
| 5090 | case BO_PtrMemI: |
| 5091 | case BO_LE: |
| 5092 | case BO_GE: |
| 5093 | case BO_EQ: |
| 5094 | case BO_NE: |
| 5095 | case BO_Cmp: |
| 5096 | case BO_AddAssign: |
| 5097 | case BO_SubAssign: |
| 5098 | case BO_AndAssign: |
| 5099 | case BO_OrAssign: |
| 5100 | case BO_XorAssign: |
| 5101 | case BO_MulAssign: |
| 5102 | case BO_DivAssign: |
| 5103 | case BO_RemAssign: |
| 5104 | case BO_ShlAssign: |
| 5105 | case BO_ShrAssign: |
| 5106 | case BO_Comma: |
| 5107 | llvm_unreachable("Unsupported atomic update operation" ); |
| 5108 | } |
| 5109 | llvm::Value *UpdateVal = Update.getScalarVal(); |
| 5110 | if (auto *IC = dyn_cast<llvm::ConstantInt>(UpdateVal)) { |
| 5111 | UpdateVal = CGF.Builder.CreateIntCast( |
| 5112 | IC, X.getAddress(CGF).getElementType(), |
| 5113 | X.getType()->hasSignedIntegerRepresentation()); |
| 5114 | } |
| 5115 | llvm::Value *Res = |
| 5116 | CGF.Builder.CreateAtomicRMW(RMWOp, X.getPointer(CGF), UpdateVal, AO); |
| 5117 | return std::make_pair(true, RValue::get(Res)); |
| 5118 | } |
| 5119 | |
| 5120 | std::pair<bool, RValue> CodeGenFunction::EmitOMPAtomicSimpleUpdateExpr( |
| 5121 | LValue X, RValue E, BinaryOperatorKind BO, bool IsXLHSInRHSPart, |
| 5122 | llvm::AtomicOrdering AO, SourceLocation Loc, |
| 5123 | const llvm::function_ref<RValue(RValue)> CommonGen) { |
| 5124 | // Update expressions are allowed to have the following forms: |
| 5125 | // x binop= expr; -> xrval + expr; |
| 5126 | // x++, ++x -> xrval + 1; |
| 5127 | // x--, --x -> xrval - 1; |
| 5128 | // x = x binop expr; -> xrval binop expr |
| 5129 | // x = expr Op x; - > expr binop xrval; |
| 5130 | auto Res = emitOMPAtomicRMW(*this, X, E, BO, AO, IsXLHSInRHSPart); |
| 5131 | if (!Res.first) { |
| 5132 | if (X.isGlobalReg()) { |
| 5133 | // Emit an update expression: 'xrval' binop 'expr' or 'expr' binop |
| 5134 | // 'xrval'. |
| 5135 | EmitStoreThroughLValue(CommonGen(EmitLoadOfLValue(X, Loc)), X); |
| 5136 | } else { |
| 5137 | // Perform compare-and-swap procedure. |
| 5138 | EmitAtomicUpdate(X, AO, CommonGen, X.getType().isVolatileQualified()); |
| 5139 | } |
| 5140 | } |
| 5141 | return Res; |
| 5142 | } |
| 5143 | |
| 5144 | static void emitOMPAtomicUpdateExpr(CodeGenFunction &CGF, |
| 5145 | llvm::AtomicOrdering AO, const Expr *X, |
| 5146 | const Expr *E, const Expr *UE, |
| 5147 | bool IsXLHSInRHSPart, SourceLocation Loc) { |
| 5148 | assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) && |
| 5149 | "Update expr in 'atomic update' must be a binary operator." ); |
| 5150 | const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts()); |
| 5151 | // Update expressions are allowed to have the following forms: |
| 5152 | // x binop= expr; -> xrval + expr; |
| 5153 | // x++, ++x -> xrval + 1; |
| 5154 | // x--, --x -> xrval - 1; |
| 5155 | // x = x binop expr; -> xrval binop expr |
| 5156 | // x = expr Op x; - > expr binop xrval; |
| 5157 | assert(X->isLValue() && "X of 'omp atomic update' is not lvalue" ); |
| 5158 | LValue XLValue = CGF.EmitLValue(X); |
| 5159 | RValue ExprRValue = CGF.EmitAnyExpr(E); |
| 5160 | const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts()); |
| 5161 | const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts()); |
| 5162 | const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS; |
| 5163 | const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS; |
| 5164 | auto &&Gen = [&CGF, UE, ExprRValue, XRValExpr, ERValExpr](RValue XRValue) { |
| 5165 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 5166 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue); |
| 5167 | return CGF.EmitAnyExpr(UE); |
| 5168 | }; |
| 5169 | (void)CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 5170 | XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen); |
| 5171 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, X); |
| 5172 | // OpenMP, 2.17.7, atomic Construct |
| 5173 | // If the write, update, or capture clause is specified and the release, |
| 5174 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 5175 | // the atomic operation is also a release flush. |
| 5176 | switch (AO) { |
| 5177 | case llvm::AtomicOrdering::Release: |
| 5178 | case llvm::AtomicOrdering::AcquireRelease: |
| 5179 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 5180 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 5181 | llvm::AtomicOrdering::Release); |
| 5182 | break; |
| 5183 | case llvm::AtomicOrdering::Acquire: |
| 5184 | case llvm::AtomicOrdering::Monotonic: |
| 5185 | break; |
| 5186 | case llvm::AtomicOrdering::NotAtomic: |
| 5187 | case llvm::AtomicOrdering::Unordered: |
| 5188 | llvm_unreachable("Unexpected ordering." ); |
| 5189 | } |
| 5190 | } |
| 5191 | |
| 5192 | static RValue convertToType(CodeGenFunction &CGF, RValue Value, |
| 5193 | QualType SourceType, QualType ResType, |
| 5194 | SourceLocation Loc) { |
| 5195 | switch (CGF.getEvaluationKind(ResType)) { |
| 5196 | case TEK_Scalar: |
| 5197 | return RValue::get( |
| 5198 | convertToScalarValue(CGF, Value, SourceType, ResType, Loc)); |
| 5199 | case TEK_Complex: { |
| 5200 | auto Res = convertToComplexValue(CGF, Value, SourceType, ResType, Loc); |
| 5201 | return RValue::getComplex(Res.first, Res.second); |
| 5202 | } |
| 5203 | case TEK_Aggregate: |
| 5204 | break; |
| 5205 | } |
| 5206 | llvm_unreachable("Must be a scalar or complex." ); |
| 5207 | } |
| 5208 | |
| 5209 | static void emitOMPAtomicCaptureExpr(CodeGenFunction &CGF, |
| 5210 | llvm::AtomicOrdering AO, |
| 5211 | bool IsPostfixUpdate, const Expr *V, |
| 5212 | const Expr *X, const Expr *E, |
| 5213 | const Expr *UE, bool IsXLHSInRHSPart, |
| 5214 | SourceLocation Loc) { |
| 5215 | assert(X->isLValue() && "X of 'omp atomic capture' is not lvalue" ); |
| 5216 | assert(V->isLValue() && "V of 'omp atomic capture' is not lvalue" ); |
| 5217 | RValue NewVVal; |
| 5218 | LValue VLValue = CGF.EmitLValue(V); |
| 5219 | LValue XLValue = CGF.EmitLValue(X); |
| 5220 | RValue ExprRValue = CGF.EmitAnyExpr(E); |
| 5221 | QualType NewVValType; |
| 5222 | if (UE) { |
| 5223 | // 'x' is updated with some additional value. |
| 5224 | assert(isa<BinaryOperator>(UE->IgnoreImpCasts()) && |
| 5225 | "Update expr in 'atomic capture' must be a binary operator." ); |
| 5226 | const auto *BOUE = cast<BinaryOperator>(UE->IgnoreImpCasts()); |
| 5227 | // Update expressions are allowed to have the following forms: |
| 5228 | // x binop= expr; -> xrval + expr; |
| 5229 | // x++, ++x -> xrval + 1; |
| 5230 | // x--, --x -> xrval - 1; |
| 5231 | // x = x binop expr; -> xrval binop expr |
| 5232 | // x = expr Op x; - > expr binop xrval; |
| 5233 | const auto *LHS = cast<OpaqueValueExpr>(BOUE->getLHS()->IgnoreImpCasts()); |
| 5234 | const auto *RHS = cast<OpaqueValueExpr>(BOUE->getRHS()->IgnoreImpCasts()); |
| 5235 | const OpaqueValueExpr *XRValExpr = IsXLHSInRHSPart ? LHS : RHS; |
| 5236 | NewVValType = XRValExpr->getType(); |
| 5237 | const OpaqueValueExpr *ERValExpr = IsXLHSInRHSPart ? RHS : LHS; |
| 5238 | auto &&Gen = [&CGF, &NewVVal, UE, ExprRValue, XRValExpr, ERValExpr, |
| 5239 | IsPostfixUpdate](RValue XRValue) { |
| 5240 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 5241 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, XRValue); |
| 5242 | RValue Res = CGF.EmitAnyExpr(UE); |
| 5243 | NewVVal = IsPostfixUpdate ? XRValue : Res; |
| 5244 | return Res; |
| 5245 | }; |
| 5246 | auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 5247 | XLValue, ExprRValue, BOUE->getOpcode(), IsXLHSInRHSPart, AO, Loc, Gen); |
| 5248 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, X); |
| 5249 | if (Res.first) { |
| 5250 | // 'atomicrmw' instruction was generated. |
| 5251 | if (IsPostfixUpdate) { |
| 5252 | // Use old value from 'atomicrmw'. |
| 5253 | NewVVal = Res.second; |
| 5254 | } else { |
| 5255 | // 'atomicrmw' does not provide new value, so evaluate it using old |
| 5256 | // value of 'x'. |
| 5257 | CodeGenFunction::OpaqueValueMapping MapExpr(CGF, ERValExpr, ExprRValue); |
| 5258 | CodeGenFunction::OpaqueValueMapping MapX(CGF, XRValExpr, Res.second); |
| 5259 | NewVVal = CGF.EmitAnyExpr(UE); |
| 5260 | } |
| 5261 | } |
| 5262 | } else { |
| 5263 | // 'x' is simply rewritten with some 'expr'. |
| 5264 | NewVValType = X->getType().getNonReferenceType(); |
| 5265 | ExprRValue = convertToType(CGF, ExprRValue, E->getType(), |
| 5266 | X->getType().getNonReferenceType(), Loc); |
| 5267 | auto &&Gen = [&NewVVal, ExprRValue](RValue XRValue) { |
| 5268 | NewVVal = XRValue; |
| 5269 | return ExprRValue; |
| 5270 | }; |
| 5271 | // Try to perform atomicrmw xchg, otherwise simple exchange. |
| 5272 | auto Res = CGF.EmitOMPAtomicSimpleUpdateExpr( |
| 5273 | XLValue, ExprRValue, /*BO=*/BO_Assign, /*IsXLHSInRHSPart=*/false, AO, |
| 5274 | Loc, Gen); |
| 5275 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, X); |
| 5276 | if (Res.first) { |
| 5277 | // 'atomicrmw' instruction was generated. |
| 5278 | NewVVal = IsPostfixUpdate ? Res.second : ExprRValue; |
| 5279 | } |
| 5280 | } |
| 5281 | // Emit post-update store to 'v' of old/new 'x' value. |
| 5282 | CGF.emitOMPSimpleStore(VLValue, NewVVal, NewVValType, Loc); |
| 5283 | CGF.CGM.getOpenMPRuntime().checkAndEmitLastprivateConditional(CGF, V); |
| 5284 | // OpenMP, 2.17.7, atomic Construct |
| 5285 | // If the write, update, or capture clause is specified and the release, |
| 5286 | // acq_rel, or seq_cst clause is specified then the strong flush on entry to |
| 5287 | // the atomic operation is also a release flush. |
| 5288 | // If the read or capture clause is specified and the acquire, acq_rel, or |
| 5289 | // seq_cst clause is specified then the strong flush on exit from the atomic |
| 5290 | // operation is also an acquire flush. |
| 5291 | switch (AO) { |
| 5292 | case llvm::AtomicOrdering::Release: |
| 5293 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 5294 | llvm::AtomicOrdering::Release); |
| 5295 | break; |
| 5296 | case llvm::AtomicOrdering::Acquire: |
| 5297 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 5298 | llvm::AtomicOrdering::Acquire); |
| 5299 | break; |
| 5300 | case llvm::AtomicOrdering::AcquireRelease: |
| 5301 | case llvm::AtomicOrdering::SequentiallyConsistent: |
| 5302 | CGF.CGM.getOpenMPRuntime().emitFlush(CGF, llvm::None, Loc, |
| 5303 | llvm::AtomicOrdering::AcquireRelease); |
| 5304 | break; |
| 5305 | case llvm::AtomicOrdering::Monotonic: |
| 5306 | break; |
| 5307 | case llvm::AtomicOrdering::NotAtomic: |
| 5308 | case llvm::AtomicOrdering::Unordered: |
| 5309 | llvm_unreachable("Unexpected ordering." ); |
| 5310 | } |
| 5311 | } |
| 5312 | |
| 5313 | static void emitOMPAtomicExpr(CodeGenFunction &CGF, OpenMPClauseKind Kind, |
| 5314 | llvm::AtomicOrdering AO, bool IsPostfixUpdate, |
| 5315 | const Expr *X, const Expr *V, const Expr *E, |
| 5316 | const Expr *UE, bool IsXLHSInRHSPart, |
| 5317 | SourceLocation Loc) { |
| 5318 | switch (Kind) { |
| 5319 | case OMPC_read: |
| 5320 | emitOMPAtomicReadExpr(CGF, AO, X, V, Loc); |
| 5321 | break; |
| 5322 | case OMPC_write: |
| 5323 | emitOMPAtomicWriteExpr(CGF, AO, X, E, Loc); |
| 5324 | break; |
| 5325 | case OMPC_unknown: |
| 5326 | case OMPC_update: |
| 5327 | emitOMPAtomicUpdateExpr(CGF, AO, X, E, UE, IsXLHSInRHSPart, Loc); |
| 5328 | break; |
| 5329 | case OMPC_capture: |
| 5330 | emitOMPAtomicCaptureExpr(CGF, AO, IsPostfixUpdate, V, X, E, UE, |
| 5331 | IsXLHSInRHSPart, Loc); |
| 5332 | break; |
| 5333 | case OMPC_if: |
| 5334 | case OMPC_final: |
| 5335 | case OMPC_num_threads: |
| 5336 | case OMPC_private: |
| 5337 | case OMPC_firstprivate: |
| 5338 | case OMPC_lastprivate: |
| 5339 | case OMPC_reduction: |
| 5340 | case OMPC_task_reduction: |
| 5341 | case OMPC_in_reduction: |
| 5342 | case OMPC_safelen: |
| 5343 | case OMPC_simdlen: |
| 5344 | case OMPC_allocator: |
| 5345 | case OMPC_allocate: |
| 5346 | case OMPC_collapse: |
| 5347 | case OMPC_default: |
| 5348 | case OMPC_seq_cst: |
| 5349 | case OMPC_acq_rel: |
| 5350 | case OMPC_acquire: |
| 5351 | case OMPC_release: |
| 5352 | case OMPC_relaxed: |
| 5353 | case OMPC_shared: |
| 5354 | case OMPC_linear: |
| 5355 | case OMPC_aligned: |
| 5356 | case OMPC_copyin: |
| 5357 | case OMPC_copyprivate: |
| 5358 | case OMPC_flush: |
| 5359 | case OMPC_depobj: |
| 5360 | case OMPC_proc_bind: |
| 5361 | case OMPC_schedule: |
| 5362 | case OMPC_ordered: |
| 5363 | case OMPC_nowait: |
| 5364 | case OMPC_untied: |
| 5365 | case OMPC_threadprivate: |
| 5366 | case OMPC_depend: |
| 5367 | case OMPC_mergeable: |
| 5368 | case OMPC_device: |
| 5369 | case OMPC_threads: |
| 5370 | case OMPC_simd: |
| 5371 | case OMPC_map: |
| 5372 | case OMPC_num_teams: |
| 5373 | case OMPC_thread_limit: |
| 5374 | case OMPC_priority: |
| 5375 | case OMPC_grainsize: |
| 5376 | case OMPC_nogroup: |
| 5377 | case OMPC_num_tasks: |
| 5378 | case OMPC_hint: |
| 5379 | case OMPC_dist_schedule: |
| 5380 | case OMPC_defaultmap: |
| 5381 | case OMPC_uniform: |
| 5382 | case OMPC_to: |
| 5383 | case OMPC_from: |
| 5384 | case OMPC_use_device_ptr: |
| 5385 | case OMPC_use_device_addr: |
| 5386 | case OMPC_is_device_ptr: |
| 5387 | case OMPC_unified_address: |
| 5388 | case OMPC_unified_shared_memory: |
| 5389 | case OMPC_reverse_offload: |
| 5390 | case OMPC_dynamic_allocators: |
| 5391 | case OMPC_atomic_default_mem_order: |
| 5392 | case OMPC_device_type: |
| 5393 | case OMPC_match: |
| 5394 | case OMPC_nontemporal: |
| 5395 | case OMPC_order: |
| 5396 | case OMPC_destroy: |
| 5397 | case OMPC_detach: |
| 5398 | case OMPC_inclusive: |
| 5399 | case OMPC_exclusive: |
| 5400 | case OMPC_uses_allocators: |
| 5401 | case OMPC_affinity: |
| 5402 | default: |
| 5403 | llvm_unreachable("Clause is not allowed in 'omp atomic'." ); |
| 5404 | } |
| 5405 | } |
| 5406 | |
| 5407 | void CodeGenFunction::EmitOMPAtomicDirective(const OMPAtomicDirective &S) { |
| 5408 | llvm::AtomicOrdering AO = llvm::AtomicOrdering::Monotonic; |
| 5409 | bool MemOrderingSpecified = false; |
| 5410 | if (S.getSingleClause<OMPSeqCstClause>()) { |
| 5411 | AO = llvm::AtomicOrdering::SequentiallyConsistent; |
| 5412 | MemOrderingSpecified = true; |
| 5413 | } else if (S.getSingleClause<OMPAcqRelClause>()) { |
| 5414 | AO = llvm::AtomicOrdering::AcquireRelease; |
| 5415 | MemOrderingSpecified = true; |
| 5416 | } else if (S.getSingleClause<OMPAcquireClause>()) { |
| 5417 | AO = llvm::AtomicOrdering::Acquire; |
| 5418 | MemOrderingSpecified = true; |
| 5419 | } else if (S.getSingleClause<OMPReleaseClause>()) { |
| 5420 | AO = llvm::AtomicOrdering::Release; |
| 5421 | MemOrderingSpecified = true; |
| 5422 | } else if (S.getSingleClause<OMPRelaxedClause>()) { |
| 5423 | AO = llvm::AtomicOrdering::Monotonic; |
| 5424 | MemOrderingSpecified = true; |
| 5425 | } |
| 5426 | OpenMPClauseKind Kind = OMPC_unknown; |
| 5427 | for (const OMPClause *C : S.clauses()) { |
| 5428 | // Find first clause (skip seq_cst|acq_rel|aqcuire|release|relaxed clause, |
| 5429 | // if it is first). |
| 5430 | if (C->getClauseKind() != OMPC_seq_cst && |
| 5431 | C->getClauseKind() != OMPC_acq_rel && |
| 5432 | C->getClauseKind() != OMPC_acquire && |
| 5433 | C->getClauseKind() != OMPC_release && |
| 5434 | C->getClauseKind() != OMPC_relaxed) { |
| 5435 | Kind = C->getClauseKind(); |
| 5436 | break; |
| 5437 | } |
| 5438 | } |
| 5439 | if (!MemOrderingSpecified) { |
| 5440 | llvm::AtomicOrdering DefaultOrder = |
| 5441 | CGM.getOpenMPRuntime().getDefaultMemoryOrdering(); |
| 5442 | if (DefaultOrder == llvm::AtomicOrdering::Monotonic || |
| 5443 | DefaultOrder == llvm::AtomicOrdering::SequentiallyConsistent || |
| 5444 | (DefaultOrder == llvm::AtomicOrdering::AcquireRelease && |
| 5445 | Kind == OMPC_capture)) { |
| 5446 | AO = DefaultOrder; |
| 5447 | } else if (DefaultOrder == llvm::AtomicOrdering::AcquireRelease) { |
| 5448 | if (Kind == OMPC_unknown || Kind == OMPC_update || Kind == OMPC_write) { |
| 5449 | AO = llvm::AtomicOrdering::Release; |
| 5450 | } else if (Kind == OMPC_read) { |
| 5451 | assert(Kind == OMPC_read && "Unexpected atomic kind." ); |
| 5452 | AO = llvm::AtomicOrdering::Acquire; |
| 5453 | } |
| 5454 | } |
| 5455 | } |
| 5456 | |
| 5457 | LexicalScope Scope(*this, S.getSourceRange()); |
| 5458 | EmitStopPoint(S.getAssociatedStmt()); |
| 5459 | emitOMPAtomicExpr(*this, Kind, AO, S.isPostfixUpdate(), S.getX(), S.getV(), |
| 5460 | S.getExpr(), S.getUpdateExpr(), S.isXLHSInRHSPart(), |
| 5461 | S.getBeginLoc()); |
| 5462 | } |
| 5463 | |
| 5464 | static void emitCommonOMPTargetDirective(CodeGenFunction &CGF, |
| 5465 | const OMPExecutableDirective &S, |
| 5466 | const RegionCodeGenTy &CodeGen) { |
| 5467 | assert(isOpenMPTargetExecutionDirective(S.getDirectiveKind())); |
| 5468 | CodeGenModule &CGM = CGF.CGM; |
| 5469 | |
| 5470 | // On device emit this construct as inlined code. |
| 5471 | if (CGM.getLangOpts().OpenMPIsDevice) { |
| 5472 | OMPLexicalScope Scope(CGF, S, OMPD_target); |
| 5473 | CGM.getOpenMPRuntime().emitInlinedDirective( |
| 5474 | CGF, OMPD_target, [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5475 | CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 5476 | }); |
| 5477 | return; |
| 5478 | } |
| 5479 | |
| 5480 | auto LPCRegion = |
| 5481 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(CGF, S); |
| 5482 | llvm::Function *Fn = nullptr; |
| 5483 | llvm::Constant *FnID = nullptr; |
| 5484 | |
| 5485 | const Expr *IfCond = nullptr; |
| 5486 | // Check for the at most one if clause associated with the target region. |
| 5487 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 5488 | if (C->getNameModifier() == OMPD_unknown || |
| 5489 | C->getNameModifier() == OMPD_target) { |
| 5490 | IfCond = C->getCondition(); |
| 5491 | break; |
| 5492 | } |
| 5493 | } |
| 5494 | |
| 5495 | // Check if we have any device clause associated with the directive. |
| 5496 | llvm::PointerIntPair<const Expr *, 2, OpenMPDeviceClauseModifier> Device( |
| 5497 | nullptr, OMPC_DEVICE_unknown); |
| 5498 | if (auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 5499 | Device.setPointerAndInt(C->getDevice(), C->getModifier()); |
| 5500 | |
| 5501 | // Check if we have an if clause whose conditional always evaluates to false |
| 5502 | // or if we do not have any targets specified. If so the target region is not |
| 5503 | // an offload entry point. |
| 5504 | bool IsOffloadEntry = true; |
| 5505 | if (IfCond) { |
| 5506 | bool Val; |
| 5507 | if (CGF.ConstantFoldsToSimpleInteger(IfCond, Val) && !Val) |
| 5508 | IsOffloadEntry = false; |
| 5509 | } |
| 5510 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 5511 | IsOffloadEntry = false; |
| 5512 | |
| 5513 | assert(CGF.CurFuncDecl && "No parent declaration for target region!" ); |
| 5514 | StringRef ParentName; |
| 5515 | // In case we have Ctors/Dtors we use the complete type variant to produce |
| 5516 | // the mangling of the device outlined kernel. |
| 5517 | if (const auto *D = dyn_cast<CXXConstructorDecl>(CGF.CurFuncDecl)) |
| 5518 | ParentName = CGM.getMangledName(GlobalDecl(D, Ctor_Complete)); |
| 5519 | else if (const auto *D = dyn_cast<CXXDestructorDecl>(CGF.CurFuncDecl)) |
| 5520 | ParentName = CGM.getMangledName(GlobalDecl(D, Dtor_Complete)); |
| 5521 | else |
| 5522 | ParentName = |
| 5523 | CGM.getMangledName(GlobalDecl(cast<FunctionDecl>(CGF.CurFuncDecl))); |
| 5524 | |
| 5525 | // Emit target region as a standalone region. |
| 5526 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction(S, ParentName, Fn, FnID, |
| 5527 | IsOffloadEntry, CodeGen); |
| 5528 | OMPLexicalScope Scope(CGF, S, OMPD_task); |
| 5529 | auto &&SizeEmitter = |
| 5530 | [IsOffloadEntry](CodeGenFunction &CGF, |
| 5531 | const OMPLoopDirective &D) -> llvm::Value * { |
| 5532 | if (IsOffloadEntry) { |
| 5533 | OMPLoopScope(CGF, D); |
| 5534 | // Emit calculation of the iterations count. |
| 5535 | llvm::Value *NumIterations = CGF.EmitScalarExpr(D.getNumIterations()); |
| 5536 | NumIterations = CGF.Builder.CreateIntCast(NumIterations, CGF.Int64Ty, |
| 5537 | /*isSigned=*/false); |
| 5538 | return NumIterations; |
| 5539 | } |
| 5540 | return nullptr; |
| 5541 | }; |
| 5542 | CGM.getOpenMPRuntime().emitTargetCall(CGF, S, Fn, FnID, IfCond, Device, |
| 5543 | SizeEmitter); |
| 5544 | } |
| 5545 | |
| 5546 | static void emitTargetRegion(CodeGenFunction &CGF, const OMPTargetDirective &S, |
| 5547 | PrePostActionTy &Action) { |
| 5548 | Action.Enter(CGF); |
| 5549 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5550 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 5551 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 5552 | (void)PrivateScope.Privatize(); |
| 5553 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 5554 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S); |
| 5555 | |
| 5556 | CGF.EmitStmt(S.getCapturedStmt(OMPD_target)->getCapturedStmt()); |
| 5557 | } |
| 5558 | |
| 5559 | void CodeGenFunction::EmitOMPTargetDeviceFunction(CodeGenModule &CGM, |
| 5560 | StringRef ParentName, |
| 5561 | const OMPTargetDirective &S) { |
| 5562 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5563 | emitTargetRegion(CGF, S, Action); |
| 5564 | }; |
| 5565 | llvm::Function *Fn; |
| 5566 | llvm::Constant *Addr; |
| 5567 | // Emit target region as a standalone region. |
| 5568 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5569 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5570 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5571 | } |
| 5572 | |
| 5573 | void CodeGenFunction::EmitOMPTargetDirective(const OMPTargetDirective &S) { |
| 5574 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5575 | emitTargetRegion(CGF, S, Action); |
| 5576 | }; |
| 5577 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5578 | } |
| 5579 | |
| 5580 | static void emitCommonOMPTeamsDirective(CodeGenFunction &CGF, |
| 5581 | const OMPExecutableDirective &S, |
| 5582 | OpenMPDirectiveKind InnermostKind, |
| 5583 | const RegionCodeGenTy &CodeGen) { |
| 5584 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_teams); |
| 5585 | llvm::Function *OutlinedFn = |
| 5586 | CGF.CGM.getOpenMPRuntime().emitTeamsOutlinedFunction( |
| 5587 | S, *CS->getCapturedDecl()->param_begin(), InnermostKind, CodeGen); |
| 5588 | |
| 5589 | const auto *NT = S.getSingleClause<OMPNumTeamsClause>(); |
| 5590 | const auto *TL = S.getSingleClause<OMPThreadLimitClause>(); |
| 5591 | if (NT || TL) { |
| 5592 | const Expr *NumTeams = NT ? NT->getNumTeams() : nullptr; |
| 5593 | const Expr *ThreadLimit = TL ? TL->getThreadLimit() : nullptr; |
| 5594 | |
| 5595 | CGF.CGM.getOpenMPRuntime().emitNumTeamsClause(CGF, NumTeams, ThreadLimit, |
| 5596 | S.getBeginLoc()); |
| 5597 | } |
| 5598 | |
| 5599 | OMPTeamsScope Scope(CGF, S); |
| 5600 | llvm::SmallVector<llvm::Value *, 16> CapturedVars; |
| 5601 | CGF.GenerateOpenMPCapturedVars(*CS, CapturedVars); |
| 5602 | CGF.CGM.getOpenMPRuntime().emitTeamsCall(CGF, S, S.getBeginLoc(), OutlinedFn, |
| 5603 | CapturedVars); |
| 5604 | } |
| 5605 | |
| 5606 | void CodeGenFunction::EmitOMPTeamsDirective(const OMPTeamsDirective &S) { |
| 5607 | // Emit teams region as a standalone region. |
| 5608 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5609 | Action.Enter(CGF); |
| 5610 | OMPPrivateScope PrivateScope(CGF); |
| 5611 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 5612 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 5613 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5614 | (void)PrivateScope.Privatize(); |
| 5615 | CGF.EmitStmt(S.getCapturedStmt(OMPD_teams)->getCapturedStmt()); |
| 5616 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5617 | }; |
| 5618 | emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen); |
| 5619 | emitPostUpdateForReductionClause(*this, S, |
| 5620 | [](CodeGenFunction &) { return nullptr; }); |
| 5621 | } |
| 5622 | |
| 5623 | static void emitTargetTeamsRegion(CodeGenFunction &CGF, PrePostActionTy &Action, |
| 5624 | const OMPTargetTeamsDirective &S) { |
| 5625 | auto *CS = S.getCapturedStmt(OMPD_teams); |
| 5626 | Action.Enter(CGF); |
| 5627 | // Emit teams region as a standalone region. |
| 5628 | auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5629 | Action.Enter(CGF); |
| 5630 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5631 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 5632 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 5633 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5634 | (void)PrivateScope.Privatize(); |
| 5635 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 5636 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S); |
| 5637 | CGF.EmitStmt(CS->getCapturedStmt()); |
| 5638 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5639 | }; |
| 5640 | emitCommonOMPTeamsDirective(CGF, S, OMPD_teams, CodeGen); |
| 5641 | emitPostUpdateForReductionClause(CGF, S, |
| 5642 | [](CodeGenFunction &) { return nullptr; }); |
| 5643 | } |
| 5644 | |
| 5645 | void CodeGenFunction::EmitOMPTargetTeamsDeviceFunction( |
| 5646 | CodeGenModule &CGM, StringRef ParentName, |
| 5647 | const OMPTargetTeamsDirective &S) { |
| 5648 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5649 | emitTargetTeamsRegion(CGF, Action, S); |
| 5650 | }; |
| 5651 | llvm::Function *Fn; |
| 5652 | llvm::Constant *Addr; |
| 5653 | // Emit target region as a standalone region. |
| 5654 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5655 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5656 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5657 | } |
| 5658 | |
| 5659 | void CodeGenFunction::EmitOMPTargetTeamsDirective( |
| 5660 | const OMPTargetTeamsDirective &S) { |
| 5661 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5662 | emitTargetTeamsRegion(CGF, Action, S); |
| 5663 | }; |
| 5664 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5665 | } |
| 5666 | |
| 5667 | static void |
| 5668 | emitTargetTeamsDistributeRegion(CodeGenFunction &CGF, PrePostActionTy &Action, |
| 5669 | const OMPTargetTeamsDistributeDirective &S) { |
| 5670 | Action.Enter(CGF); |
| 5671 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5672 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 5673 | }; |
| 5674 | |
| 5675 | // Emit teams region as a standalone region. |
| 5676 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5677 | PrePostActionTy &Action) { |
| 5678 | Action.Enter(CGF); |
| 5679 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5680 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5681 | (void)PrivateScope.Privatize(); |
| 5682 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute, |
| 5683 | CodeGenDistribute); |
| 5684 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5685 | }; |
| 5686 | emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute, CodeGen); |
| 5687 | emitPostUpdateForReductionClause(CGF, S, |
| 5688 | [](CodeGenFunction &) { return nullptr; }); |
| 5689 | } |
| 5690 | |
| 5691 | void CodeGenFunction::EmitOMPTargetTeamsDistributeDeviceFunction( |
| 5692 | CodeGenModule &CGM, StringRef ParentName, |
| 5693 | const OMPTargetTeamsDistributeDirective &S) { |
| 5694 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5695 | emitTargetTeamsDistributeRegion(CGF, Action, S); |
| 5696 | }; |
| 5697 | llvm::Function *Fn; |
| 5698 | llvm::Constant *Addr; |
| 5699 | // Emit target region as a standalone region. |
| 5700 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5701 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5702 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5703 | } |
| 5704 | |
| 5705 | void CodeGenFunction::EmitOMPTargetTeamsDistributeDirective( |
| 5706 | const OMPTargetTeamsDistributeDirective &S) { |
| 5707 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5708 | emitTargetTeamsDistributeRegion(CGF, Action, S); |
| 5709 | }; |
| 5710 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5711 | } |
| 5712 | |
| 5713 | static void emitTargetTeamsDistributeSimdRegion( |
| 5714 | CodeGenFunction &CGF, PrePostActionTy &Action, |
| 5715 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 5716 | Action.Enter(CGF); |
| 5717 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5718 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 5719 | }; |
| 5720 | |
| 5721 | // Emit teams region as a standalone region. |
| 5722 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5723 | PrePostActionTy &Action) { |
| 5724 | Action.Enter(CGF); |
| 5725 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5726 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5727 | (void)PrivateScope.Privatize(); |
| 5728 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute, |
| 5729 | CodeGenDistribute); |
| 5730 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5731 | }; |
| 5732 | emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_simd, CodeGen); |
| 5733 | emitPostUpdateForReductionClause(CGF, S, |
| 5734 | [](CodeGenFunction &) { return nullptr; }); |
| 5735 | } |
| 5736 | |
| 5737 | void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDeviceFunction( |
| 5738 | CodeGenModule &CGM, StringRef ParentName, |
| 5739 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 5740 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5741 | emitTargetTeamsDistributeSimdRegion(CGF, Action, S); |
| 5742 | }; |
| 5743 | llvm::Function *Fn; |
| 5744 | llvm::Constant *Addr; |
| 5745 | // Emit target region as a standalone region. |
| 5746 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5747 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5748 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5749 | } |
| 5750 | |
| 5751 | void CodeGenFunction::EmitOMPTargetTeamsDistributeSimdDirective( |
| 5752 | const OMPTargetTeamsDistributeSimdDirective &S) { |
| 5753 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5754 | emitTargetTeamsDistributeSimdRegion(CGF, Action, S); |
| 5755 | }; |
| 5756 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5757 | } |
| 5758 | |
| 5759 | void CodeGenFunction::EmitOMPTeamsDistributeDirective( |
| 5760 | const OMPTeamsDistributeDirective &S) { |
| 5761 | |
| 5762 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5763 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 5764 | }; |
| 5765 | |
| 5766 | // Emit teams region as a standalone region. |
| 5767 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5768 | PrePostActionTy &Action) { |
| 5769 | Action.Enter(CGF); |
| 5770 | OMPPrivateScope PrivateScope(CGF); |
| 5771 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5772 | (void)PrivateScope.Privatize(); |
| 5773 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute, |
| 5774 | CodeGenDistribute); |
| 5775 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5776 | }; |
| 5777 | emitCommonOMPTeamsDirective(*this, S, OMPD_distribute, CodeGen); |
| 5778 | emitPostUpdateForReductionClause(*this, S, |
| 5779 | [](CodeGenFunction &) { return nullptr; }); |
| 5780 | } |
| 5781 | |
| 5782 | void CodeGenFunction::EmitOMPTeamsDistributeSimdDirective( |
| 5783 | const OMPTeamsDistributeSimdDirective &S) { |
| 5784 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5785 | CGF.EmitOMPDistributeLoop(S, emitOMPLoopBodyWithStopPoint, S.getInc()); |
| 5786 | }; |
| 5787 | |
| 5788 | // Emit teams region as a standalone region. |
| 5789 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5790 | PrePostActionTy &Action) { |
| 5791 | Action.Enter(CGF); |
| 5792 | OMPPrivateScope PrivateScope(CGF); |
| 5793 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5794 | (void)PrivateScope.Privatize(); |
| 5795 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_simd, |
| 5796 | CodeGenDistribute); |
| 5797 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5798 | }; |
| 5799 | emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_simd, CodeGen); |
| 5800 | emitPostUpdateForReductionClause(*this, S, |
| 5801 | [](CodeGenFunction &) { return nullptr; }); |
| 5802 | } |
| 5803 | |
| 5804 | void CodeGenFunction::EmitOMPTeamsDistributeParallelForDirective( |
| 5805 | const OMPTeamsDistributeParallelForDirective &S) { |
| 5806 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5807 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 5808 | S.getDistInc()); |
| 5809 | }; |
| 5810 | |
| 5811 | // Emit teams region as a standalone region. |
| 5812 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5813 | PrePostActionTy &Action) { |
| 5814 | Action.Enter(CGF); |
| 5815 | OMPPrivateScope PrivateScope(CGF); |
| 5816 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5817 | (void)PrivateScope.Privatize(); |
| 5818 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_distribute, |
| 5819 | CodeGenDistribute); |
| 5820 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5821 | }; |
| 5822 | emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for, CodeGen); |
| 5823 | emitPostUpdateForReductionClause(*this, S, |
| 5824 | [](CodeGenFunction &) { return nullptr; }); |
| 5825 | } |
| 5826 | |
| 5827 | void CodeGenFunction::EmitOMPTeamsDistributeParallelForSimdDirective( |
| 5828 | const OMPTeamsDistributeParallelForSimdDirective &S) { |
| 5829 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5830 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 5831 | S.getDistInc()); |
| 5832 | }; |
| 5833 | |
| 5834 | // Emit teams region as a standalone region. |
| 5835 | auto &&CodeGen = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5836 | PrePostActionTy &Action) { |
| 5837 | Action.Enter(CGF); |
| 5838 | OMPPrivateScope PrivateScope(CGF); |
| 5839 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5840 | (void)PrivateScope.Privatize(); |
| 5841 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 5842 | CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false); |
| 5843 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5844 | }; |
| 5845 | emitCommonOMPTeamsDirective(*this, S, OMPD_distribute_parallel_for_simd, |
| 5846 | CodeGen); |
| 5847 | emitPostUpdateForReductionClause(*this, S, |
| 5848 | [](CodeGenFunction &) { return nullptr; }); |
| 5849 | } |
| 5850 | |
| 5851 | static void emitTargetTeamsDistributeParallelForRegion( |
| 5852 | CodeGenFunction &CGF, const OMPTargetTeamsDistributeParallelForDirective &S, |
| 5853 | PrePostActionTy &Action) { |
| 5854 | Action.Enter(CGF); |
| 5855 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5856 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 5857 | S.getDistInc()); |
| 5858 | }; |
| 5859 | |
| 5860 | // Emit teams region as a standalone region. |
| 5861 | auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5862 | PrePostActionTy &Action) { |
| 5863 | Action.Enter(CGF); |
| 5864 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5865 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5866 | (void)PrivateScope.Privatize(); |
| 5867 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 5868 | CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false); |
| 5869 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5870 | }; |
| 5871 | |
| 5872 | emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for, |
| 5873 | CodeGenTeams); |
| 5874 | emitPostUpdateForReductionClause(CGF, S, |
| 5875 | [](CodeGenFunction &) { return nullptr; }); |
| 5876 | } |
| 5877 | |
| 5878 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDeviceFunction( |
| 5879 | CodeGenModule &CGM, StringRef ParentName, |
| 5880 | const OMPTargetTeamsDistributeParallelForDirective &S) { |
| 5881 | // Emit SPMD target teams distribute parallel for region as a standalone |
| 5882 | // region. |
| 5883 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5884 | emitTargetTeamsDistributeParallelForRegion(CGF, S, Action); |
| 5885 | }; |
| 5886 | llvm::Function *Fn; |
| 5887 | llvm::Constant *Addr; |
| 5888 | // Emit target region as a standalone region. |
| 5889 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5890 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5891 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5892 | } |
| 5893 | |
| 5894 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForDirective( |
| 5895 | const OMPTargetTeamsDistributeParallelForDirective &S) { |
| 5896 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5897 | emitTargetTeamsDistributeParallelForRegion(CGF, S, Action); |
| 5898 | }; |
| 5899 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5900 | } |
| 5901 | |
| 5902 | static void emitTargetTeamsDistributeParallelForSimdRegion( |
| 5903 | CodeGenFunction &CGF, |
| 5904 | const OMPTargetTeamsDistributeParallelForSimdDirective &S, |
| 5905 | PrePostActionTy &Action) { |
| 5906 | Action.Enter(CGF); |
| 5907 | auto &&CodeGenDistribute = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 5908 | CGF.EmitOMPDistributeLoop(S, emitInnerParallelForWhenCombined, |
| 5909 | S.getDistInc()); |
| 5910 | }; |
| 5911 | |
| 5912 | // Emit teams region as a standalone region. |
| 5913 | auto &&CodeGenTeams = [&S, &CodeGenDistribute](CodeGenFunction &CGF, |
| 5914 | PrePostActionTy &Action) { |
| 5915 | Action.Enter(CGF); |
| 5916 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 5917 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 5918 | (void)PrivateScope.Privatize(); |
| 5919 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective( |
| 5920 | CGF, OMPD_distribute, CodeGenDistribute, /*HasCancel=*/false); |
| 5921 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_teams); |
| 5922 | }; |
| 5923 | |
| 5924 | emitCommonOMPTeamsDirective(CGF, S, OMPD_distribute_parallel_for_simd, |
| 5925 | CodeGenTeams); |
| 5926 | emitPostUpdateForReductionClause(CGF, S, |
| 5927 | [](CodeGenFunction &) { return nullptr; }); |
| 5928 | } |
| 5929 | |
| 5930 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDeviceFunction( |
| 5931 | CodeGenModule &CGM, StringRef ParentName, |
| 5932 | const OMPTargetTeamsDistributeParallelForSimdDirective &S) { |
| 5933 | // Emit SPMD target teams distribute parallel for simd region as a standalone |
| 5934 | // region. |
| 5935 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5936 | emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action); |
| 5937 | }; |
| 5938 | llvm::Function *Fn; |
| 5939 | llvm::Constant *Addr; |
| 5940 | // Emit target region as a standalone region. |
| 5941 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 5942 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 5943 | assert(Fn && Addr && "Target device function emission failed." ); |
| 5944 | } |
| 5945 | |
| 5946 | void CodeGenFunction::EmitOMPTargetTeamsDistributeParallelForSimdDirective( |
| 5947 | const OMPTargetTeamsDistributeParallelForSimdDirective &S) { |
| 5948 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 5949 | emitTargetTeamsDistributeParallelForSimdRegion(CGF, S, Action); |
| 5950 | }; |
| 5951 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 5952 | } |
| 5953 | |
| 5954 | void CodeGenFunction::EmitOMPCancellationPointDirective( |
| 5955 | const OMPCancellationPointDirective &S) { |
| 5956 | CGM.getOpenMPRuntime().emitCancellationPointCall(*this, S.getBeginLoc(), |
| 5957 | S.getCancelRegion()); |
| 5958 | } |
| 5959 | |
| 5960 | void CodeGenFunction::EmitOMPCancelDirective(const OMPCancelDirective &S) { |
| 5961 | const Expr *IfCond = nullptr; |
| 5962 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 5963 | if (C->getNameModifier() == OMPD_unknown || |
| 5964 | C->getNameModifier() == OMPD_cancel) { |
| 5965 | IfCond = C->getCondition(); |
| 5966 | break; |
| 5967 | } |
| 5968 | } |
| 5969 | if (CGM.getLangOpts().OpenMPIRBuilder) { |
| 5970 | llvm::OpenMPIRBuilder &OMPBuilder = CGM.getOpenMPRuntime().getOMPBuilder(); |
| 5971 | // TODO: This check is necessary as we only generate `omp parallel` through |
| 5972 | // the OpenMPIRBuilder for now. |
| 5973 | if (S.getCancelRegion() == OMPD_parallel) { |
| 5974 | llvm::Value *IfCondition = nullptr; |
| 5975 | if (IfCond) |
| 5976 | IfCondition = EmitScalarExpr(IfCond, |
| 5977 | /*IgnoreResultAssign=*/true); |
| 5978 | return Builder.restoreIP( |
| 5979 | OMPBuilder.createCancel(Builder, IfCondition, S.getCancelRegion())); |
| 5980 | } |
| 5981 | } |
| 5982 | |
| 5983 | CGM.getOpenMPRuntime().emitCancelCall(*this, S.getBeginLoc(), IfCond, |
| 5984 | S.getCancelRegion()); |
| 5985 | } |
| 5986 | |
| 5987 | CodeGenFunction::JumpDest |
| 5988 | CodeGenFunction::getOMPCancelDestination(OpenMPDirectiveKind Kind) { |
| 5989 | if (Kind == OMPD_parallel || Kind == OMPD_task || |
| 5990 | Kind == OMPD_target_parallel || Kind == OMPD_taskloop || |
| 5991 | Kind == OMPD_master_taskloop || Kind == OMPD_parallel_master_taskloop) |
| 5992 | return ReturnBlock; |
| 5993 | assert(Kind == OMPD_for || Kind == OMPD_section || Kind == OMPD_sections || |
| 5994 | Kind == OMPD_parallel_sections || Kind == OMPD_parallel_for || |
| 5995 | Kind == OMPD_distribute_parallel_for || |
| 5996 | Kind == OMPD_target_parallel_for || |
| 5997 | Kind == OMPD_teams_distribute_parallel_for || |
| 5998 | Kind == OMPD_target_teams_distribute_parallel_for); |
| 5999 | return OMPCancelStack.getExitBlock(); |
| 6000 | } |
| 6001 | |
| 6002 | void CodeGenFunction::EmitOMPUseDevicePtrClause( |
| 6003 | const OMPUseDevicePtrClause &C, OMPPrivateScope &PrivateScope, |
| 6004 | const llvm::DenseMap<const ValueDecl *, Address> &CaptureDeviceAddrMap) { |
| 6005 | auto OrigVarIt = C.varlist_begin(); |
| 6006 | auto InitIt = C.inits().begin(); |
| 6007 | for (const Expr *PvtVarIt : C.private_copies()) { |
| 6008 | const auto *OrigVD = cast<VarDecl>(cast<DeclRefExpr>(*OrigVarIt)->getDecl()); |
| 6009 | const auto *InitVD = cast<VarDecl>(cast<DeclRefExpr>(*InitIt)->getDecl()); |
| 6010 | const auto *PvtVD = cast<VarDecl>(cast<DeclRefExpr>(PvtVarIt)->getDecl()); |
| 6011 | |
| 6012 | // In order to identify the right initializer we need to match the |
| 6013 | // declaration used by the mapping logic. In some cases we may get |
| 6014 | // OMPCapturedExprDecl that refers to the original declaration. |
| 6015 | const ValueDecl *MatchingVD = OrigVD; |
| 6016 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(MatchingVD)) { |
| 6017 | // OMPCapturedExprDecl are used to privative fields of the current |
| 6018 | // structure. |
| 6019 | const auto *ME = cast<MemberExpr>(OED->getInit()); |
| 6020 | assert(isa<CXXThisExpr>(ME->getBase()) && |
| 6021 | "Base should be the current struct!" ); |
| 6022 | MatchingVD = ME->getMemberDecl(); |
| 6023 | } |
| 6024 | |
| 6025 | // If we don't have information about the current list item, move on to |
| 6026 | // the next one. |
| 6027 | auto InitAddrIt = CaptureDeviceAddrMap.find(MatchingVD); |
| 6028 | if (InitAddrIt == CaptureDeviceAddrMap.end()) |
| 6029 | continue; |
| 6030 | |
| 6031 | bool IsRegistered = PrivateScope.addPrivate(OrigVD, [this, OrigVD, |
| 6032 | InitAddrIt, InitVD, |
| 6033 | PvtVD]() { |
| 6034 | // Initialize the temporary initialization variable with the address we |
| 6035 | // get from the runtime library. We have to cast the source address |
| 6036 | // because it is always a void *. References are materialized in the |
| 6037 | // privatization scope, so the initialization here disregards the fact |
| 6038 | // the original variable is a reference. |
| 6039 | QualType AddrQTy = |
| 6040 | getContext().getPointerType(OrigVD->getType().getNonReferenceType()); |
| 6041 | llvm::Type *AddrTy = ConvertTypeForMem(AddrQTy); |
| 6042 | Address InitAddr = Builder.CreateBitCast(InitAddrIt->second, AddrTy); |
| 6043 | setAddrOfLocalVar(InitVD, InitAddr); |
| 6044 | |
| 6045 | // Emit private declaration, it will be initialized by the value we |
| 6046 | // declaration we just added to the local declarations map. |
| 6047 | EmitDecl(*PvtVD); |
| 6048 | |
| 6049 | // The initialization variables reached its purpose in the emission |
| 6050 | // of the previous declaration, so we don't need it anymore. |
| 6051 | LocalDeclMap.erase(InitVD); |
| 6052 | |
| 6053 | // Return the address of the private variable. |
| 6054 | return GetAddrOfLocalVar(PvtVD); |
| 6055 | }); |
| 6056 | assert(IsRegistered && "firstprivate var already registered as private" ); |
| 6057 | // Silence the warning about unused variable. |
| 6058 | (void)IsRegistered; |
| 6059 | |
| 6060 | ++OrigVarIt; |
| 6061 | ++InitIt; |
| 6062 | } |
| 6063 | } |
| 6064 | |
| 6065 | static const VarDecl *getBaseDecl(const Expr *Ref) { |
| 6066 | const Expr *Base = Ref->IgnoreParenImpCasts(); |
| 6067 | while (const auto *OASE = dyn_cast<OMPArraySectionExpr>(Base)) |
| 6068 | Base = OASE->getBase()->IgnoreParenImpCasts(); |
| 6069 | while (const auto *ASE = dyn_cast<ArraySubscriptExpr>(Base)) |
| 6070 | Base = ASE->getBase()->IgnoreParenImpCasts(); |
| 6071 | return cast<VarDecl>(cast<DeclRefExpr>(Base)->getDecl()); |
| 6072 | } |
| 6073 | |
| 6074 | void CodeGenFunction::EmitOMPUseDeviceAddrClause( |
| 6075 | const OMPUseDeviceAddrClause &C, OMPPrivateScope &PrivateScope, |
| 6076 | const llvm::DenseMap<const ValueDecl *, Address> &CaptureDeviceAddrMap) { |
| 6077 | llvm::SmallDenseSet<CanonicalDeclPtr<const Decl>, 4> Processed; |
| 6078 | for (const Expr *Ref : C.varlists()) { |
| 6079 | const VarDecl *OrigVD = getBaseDecl(Ref); |
| 6080 | if (!Processed.insert(OrigVD).second) |
| 6081 | continue; |
| 6082 | // In order to identify the right initializer we need to match the |
| 6083 | // declaration used by the mapping logic. In some cases we may get |
| 6084 | // OMPCapturedExprDecl that refers to the original declaration. |
| 6085 | const ValueDecl *MatchingVD = OrigVD; |
| 6086 | if (const auto *OED = dyn_cast<OMPCapturedExprDecl>(MatchingVD)) { |
| 6087 | // OMPCapturedExprDecl are used to privative fields of the current |
| 6088 | // structure. |
| 6089 | const auto *ME = cast<MemberExpr>(OED->getInit()); |
| 6090 | assert(isa<CXXThisExpr>(ME->getBase()) && |
| 6091 | "Base should be the current struct!" ); |
| 6092 | MatchingVD = ME->getMemberDecl(); |
| 6093 | } |
| 6094 | |
| 6095 | // If we don't have information about the current list item, move on to |
| 6096 | // the next one. |
| 6097 | auto InitAddrIt = CaptureDeviceAddrMap.find(MatchingVD); |
| 6098 | if (InitAddrIt == CaptureDeviceAddrMap.end()) |
| 6099 | continue; |
| 6100 | |
| 6101 | Address PrivAddr = InitAddrIt->getSecond(); |
| 6102 | // For declrefs and variable length array need to load the pointer for |
| 6103 | // correct mapping, since the pointer to the data was passed to the runtime. |
| 6104 | if (isa<DeclRefExpr>(Ref->IgnoreParenImpCasts()) || |
| 6105 | MatchingVD->getType()->isArrayType()) |
| 6106 | PrivAddr = |
| 6107 | EmitLoadOfPointer(PrivAddr, getContext() |
| 6108 | .getPointerType(OrigVD->getType()) |
| 6109 | ->castAs<PointerType>()); |
| 6110 | llvm::Type *RealTy = |
| 6111 | ConvertTypeForMem(OrigVD->getType().getNonReferenceType()) |
| 6112 | ->getPointerTo(); |
| 6113 | PrivAddr = Builder.CreatePointerBitCastOrAddrSpaceCast(PrivAddr, RealTy); |
| 6114 | |
| 6115 | (void)PrivateScope.addPrivate(OrigVD, [PrivAddr]() { return PrivAddr; }); |
| 6116 | } |
| 6117 | } |
| 6118 | |
| 6119 | // Generate the instructions for '#pragma omp target data' directive. |
| 6120 | void CodeGenFunction::EmitOMPTargetDataDirective( |
| 6121 | const OMPTargetDataDirective &S) { |
| 6122 | CGOpenMPRuntime::TargetDataInfo Info(/*RequiresDevicePointerInfo=*/true, |
| 6123 | /*SeparateBeginEndCalls=*/true); |
| 6124 | |
| 6125 | // Create a pre/post action to signal the privatization of the device pointer. |
| 6126 | // This action can be replaced by the OpenMP runtime code generation to |
| 6127 | // deactivate privatization. |
| 6128 | bool PrivatizeDevicePointers = false; |
| 6129 | class DevicePointerPrivActionTy : public PrePostActionTy { |
| 6130 | bool &PrivatizeDevicePointers; |
| 6131 | |
| 6132 | public: |
| 6133 | explicit DevicePointerPrivActionTy(bool &PrivatizeDevicePointers) |
| 6134 | : PrePostActionTy(), PrivatizeDevicePointers(PrivatizeDevicePointers) {} |
| 6135 | void Enter(CodeGenFunction &CGF) override { |
| 6136 | PrivatizeDevicePointers = true; |
| 6137 | } |
| 6138 | }; |
| 6139 | DevicePointerPrivActionTy PrivAction(PrivatizeDevicePointers); |
| 6140 | |
| 6141 | auto &&CodeGen = [&S, &Info, &PrivatizeDevicePointers]( |
| 6142 | CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6143 | auto &&InnermostCodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6144 | CGF.EmitStmt(S.getInnermostCapturedStmt()->getCapturedStmt()); |
| 6145 | }; |
| 6146 | |
| 6147 | // Codegen that selects whether to generate the privatization code or not. |
| 6148 | auto &&PrivCodeGen = [&S, &Info, &PrivatizeDevicePointers, |
| 6149 | &InnermostCodeGen](CodeGenFunction &CGF, |
| 6150 | PrePostActionTy &Action) { |
| 6151 | RegionCodeGenTy RCG(InnermostCodeGen); |
| 6152 | PrivatizeDevicePointers = false; |
| 6153 | |
| 6154 | // Call the pre-action to change the status of PrivatizeDevicePointers if |
| 6155 | // needed. |
| 6156 | Action.Enter(CGF); |
| 6157 | |
| 6158 | if (PrivatizeDevicePointers) { |
| 6159 | OMPPrivateScope PrivateScope(CGF); |
| 6160 | // Emit all instances of the use_device_ptr clause. |
| 6161 | for (const auto *C : S.getClausesOfKind<OMPUseDevicePtrClause>()) |
| 6162 | CGF.EmitOMPUseDevicePtrClause(*C, PrivateScope, |
| 6163 | Info.CaptureDeviceAddrMap); |
| 6164 | for (const auto *C : S.getClausesOfKind<OMPUseDeviceAddrClause>()) |
| 6165 | CGF.EmitOMPUseDeviceAddrClause(*C, PrivateScope, |
| 6166 | Info.CaptureDeviceAddrMap); |
| 6167 | (void)PrivateScope.Privatize(); |
| 6168 | RCG(CGF); |
| 6169 | } else { |
| 6170 | OMPLexicalScope Scope(CGF, S, OMPD_unknown); |
| 6171 | RCG(CGF); |
| 6172 | } |
| 6173 | }; |
| 6174 | |
| 6175 | // Forward the provided action to the privatization codegen. |
| 6176 | RegionCodeGenTy PrivRCG(PrivCodeGen); |
| 6177 | PrivRCG.setAction(Action); |
| 6178 | |
| 6179 | // Notwithstanding the body of the region is emitted as inlined directive, |
| 6180 | // we don't use an inline scope as changes in the references inside the |
| 6181 | // region are expected to be visible outside, so we do not privative them. |
| 6182 | OMPLexicalScope Scope(CGF, S); |
| 6183 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_target_data, |
| 6184 | PrivRCG); |
| 6185 | }; |
| 6186 | |
| 6187 | RegionCodeGenTy RCG(CodeGen); |
| 6188 | |
| 6189 | // If we don't have target devices, don't bother emitting the data mapping |
| 6190 | // code. |
| 6191 | if (CGM.getLangOpts().OMPTargetTriples.empty()) { |
| 6192 | RCG(*this); |
| 6193 | return; |
| 6194 | } |
| 6195 | |
| 6196 | // Check if we have any if clause associated with the directive. |
| 6197 | const Expr *IfCond = nullptr; |
| 6198 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 6199 | IfCond = C->getCondition(); |
| 6200 | |
| 6201 | // Check if we have any device clause associated with the directive. |
| 6202 | const Expr *Device = nullptr; |
| 6203 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 6204 | Device = C->getDevice(); |
| 6205 | |
| 6206 | // Set the action to signal privatization of device pointers. |
| 6207 | RCG.setAction(PrivAction); |
| 6208 | |
| 6209 | // Emit region code. |
| 6210 | CGM.getOpenMPRuntime().emitTargetDataCalls(*this, S, IfCond, Device, RCG, |
| 6211 | Info); |
| 6212 | } |
| 6213 | |
| 6214 | void CodeGenFunction::EmitOMPTargetEnterDataDirective( |
| 6215 | const OMPTargetEnterDataDirective &S) { |
| 6216 | // If we don't have target devices, don't bother emitting the data mapping |
| 6217 | // code. |
| 6218 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 6219 | return; |
| 6220 | |
| 6221 | // Check if we have any if clause associated with the directive. |
| 6222 | const Expr *IfCond = nullptr; |
| 6223 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 6224 | IfCond = C->getCondition(); |
| 6225 | |
| 6226 | // Check if we have any device clause associated with the directive. |
| 6227 | const Expr *Device = nullptr; |
| 6228 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 6229 | Device = C->getDevice(); |
| 6230 | |
| 6231 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 6232 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device); |
| 6233 | } |
| 6234 | |
| 6235 | void CodeGenFunction::EmitOMPTargetExitDataDirective( |
| 6236 | const OMPTargetExitDataDirective &S) { |
| 6237 | // If we don't have target devices, don't bother emitting the data mapping |
| 6238 | // code. |
| 6239 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 6240 | return; |
| 6241 | |
| 6242 | // Check if we have any if clause associated with the directive. |
| 6243 | const Expr *IfCond = nullptr; |
| 6244 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 6245 | IfCond = C->getCondition(); |
| 6246 | |
| 6247 | // Check if we have any device clause associated with the directive. |
| 6248 | const Expr *Device = nullptr; |
| 6249 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 6250 | Device = C->getDevice(); |
| 6251 | |
| 6252 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 6253 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device); |
| 6254 | } |
| 6255 | |
| 6256 | static void emitTargetParallelRegion(CodeGenFunction &CGF, |
| 6257 | const OMPTargetParallelDirective &S, |
| 6258 | PrePostActionTy &Action) { |
| 6259 | // Get the captured statement associated with the 'parallel' region. |
| 6260 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_parallel); |
| 6261 | Action.Enter(CGF); |
| 6262 | auto &&CodeGen = [&S, CS](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6263 | Action.Enter(CGF); |
| 6264 | CodeGenFunction::OMPPrivateScope PrivateScope(CGF); |
| 6265 | (void)CGF.EmitOMPFirstprivateClause(S, PrivateScope); |
| 6266 | CGF.EmitOMPPrivateClause(S, PrivateScope); |
| 6267 | CGF.EmitOMPReductionClauseInit(S, PrivateScope); |
| 6268 | (void)PrivateScope.Privatize(); |
| 6269 | if (isOpenMPTargetExecutionDirective(S.getDirectiveKind())) |
| 6270 | CGF.CGM.getOpenMPRuntime().adjustTargetSpecificDataForLambdas(CGF, S); |
| 6271 | // TODO: Add support for clauses. |
| 6272 | CGF.EmitStmt(CS->getCapturedStmt()); |
| 6273 | CGF.EmitOMPReductionClauseFinal(S, /*ReductionKind=*/OMPD_parallel); |
| 6274 | }; |
| 6275 | emitCommonOMPParallelDirective(CGF, S, OMPD_parallel, CodeGen, |
| 6276 | emitEmptyBoundParameters); |
| 6277 | emitPostUpdateForReductionClause(CGF, S, |
| 6278 | [](CodeGenFunction &) { return nullptr; }); |
| 6279 | } |
| 6280 | |
| 6281 | void CodeGenFunction::EmitOMPTargetParallelDeviceFunction( |
| 6282 | CodeGenModule &CGM, StringRef ParentName, |
| 6283 | const OMPTargetParallelDirective &S) { |
| 6284 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6285 | emitTargetParallelRegion(CGF, S, Action); |
| 6286 | }; |
| 6287 | llvm::Function *Fn; |
| 6288 | llvm::Constant *Addr; |
| 6289 | // Emit target region as a standalone region. |
| 6290 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 6291 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 6292 | assert(Fn && Addr && "Target device function emission failed." ); |
| 6293 | } |
| 6294 | |
| 6295 | void CodeGenFunction::EmitOMPTargetParallelDirective( |
| 6296 | const OMPTargetParallelDirective &S) { |
| 6297 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6298 | emitTargetParallelRegion(CGF, S, Action); |
| 6299 | }; |
| 6300 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 6301 | } |
| 6302 | |
| 6303 | static void emitTargetParallelForRegion(CodeGenFunction &CGF, |
| 6304 | const OMPTargetParallelForDirective &S, |
| 6305 | PrePostActionTy &Action) { |
| 6306 | Action.Enter(CGF); |
| 6307 | // Emit directive as a combined directive that consists of two implicit |
| 6308 | // directives: 'parallel' with 'for' directive. |
| 6309 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6310 | Action.Enter(CGF); |
| 6311 | CodeGenFunction::OMPCancelStackRAII CancelRegion( |
| 6312 | CGF, OMPD_target_parallel_for, S.hasCancel()); |
| 6313 | CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds, |
| 6314 | emitDispatchForLoopBounds); |
| 6315 | }; |
| 6316 | emitCommonOMPParallelDirective(CGF, S, OMPD_for, CodeGen, |
| 6317 | emitEmptyBoundParameters); |
| 6318 | } |
| 6319 | |
| 6320 | void CodeGenFunction::EmitOMPTargetParallelForDeviceFunction( |
| 6321 | CodeGenModule &CGM, StringRef ParentName, |
| 6322 | const OMPTargetParallelForDirective &S) { |
| 6323 | // Emit SPMD target parallel for region as a standalone region. |
| 6324 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6325 | emitTargetParallelForRegion(CGF, S, Action); |
| 6326 | }; |
| 6327 | llvm::Function *Fn; |
| 6328 | llvm::Constant *Addr; |
| 6329 | // Emit target region as a standalone region. |
| 6330 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 6331 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 6332 | assert(Fn && Addr && "Target device function emission failed." ); |
| 6333 | } |
| 6334 | |
| 6335 | void CodeGenFunction::EmitOMPTargetParallelForDirective( |
| 6336 | const OMPTargetParallelForDirective &S) { |
| 6337 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6338 | emitTargetParallelForRegion(CGF, S, Action); |
| 6339 | }; |
| 6340 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 6341 | } |
| 6342 | |
| 6343 | static void |
| 6344 | emitTargetParallelForSimdRegion(CodeGenFunction &CGF, |
| 6345 | const OMPTargetParallelForSimdDirective &S, |
| 6346 | PrePostActionTy &Action) { |
| 6347 | Action.Enter(CGF); |
| 6348 | // Emit directive as a combined directive that consists of two implicit |
| 6349 | // directives: 'parallel' with 'for' directive. |
| 6350 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6351 | Action.Enter(CGF); |
| 6352 | CGF.EmitOMPWorksharingLoop(S, S.getEnsureUpperBound(), emitForLoopBounds, |
| 6353 | emitDispatchForLoopBounds); |
| 6354 | }; |
| 6355 | emitCommonOMPParallelDirective(CGF, S, OMPD_simd, CodeGen, |
| 6356 | emitEmptyBoundParameters); |
| 6357 | } |
| 6358 | |
| 6359 | void CodeGenFunction::EmitOMPTargetParallelForSimdDeviceFunction( |
| 6360 | CodeGenModule &CGM, StringRef ParentName, |
| 6361 | const OMPTargetParallelForSimdDirective &S) { |
| 6362 | // Emit SPMD target parallel for region as a standalone region. |
| 6363 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6364 | emitTargetParallelForSimdRegion(CGF, S, Action); |
| 6365 | }; |
| 6366 | llvm::Function *Fn; |
| 6367 | llvm::Constant *Addr; |
| 6368 | // Emit target region as a standalone region. |
| 6369 | CGM.getOpenMPRuntime().emitTargetOutlinedFunction( |
| 6370 | S, ParentName, Fn, Addr, /*IsOffloadEntry=*/true, CodeGen); |
| 6371 | assert(Fn && Addr && "Target device function emission failed." ); |
| 6372 | } |
| 6373 | |
| 6374 | void CodeGenFunction::EmitOMPTargetParallelForSimdDirective( |
| 6375 | const OMPTargetParallelForSimdDirective &S) { |
| 6376 | auto &&CodeGen = [&S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6377 | emitTargetParallelForSimdRegion(CGF, S, Action); |
| 6378 | }; |
| 6379 | emitCommonOMPTargetDirective(*this, S, CodeGen); |
| 6380 | } |
| 6381 | |
| 6382 | /// Emit a helper variable and return corresponding lvalue. |
| 6383 | static void mapParam(CodeGenFunction &CGF, const DeclRefExpr *Helper, |
| 6384 | const ImplicitParamDecl *PVD, |
| 6385 | CodeGenFunction::OMPPrivateScope &Privates) { |
| 6386 | const auto *VDecl = cast<VarDecl>(Helper->getDecl()); |
| 6387 | Privates.addPrivate(VDecl, |
| 6388 | [&CGF, PVD]() { return CGF.GetAddrOfLocalVar(PVD); }); |
| 6389 | } |
| 6390 | |
| 6391 | void CodeGenFunction::EmitOMPTaskLoopBasedDirective(const OMPLoopDirective &S) { |
| 6392 | assert(isOpenMPTaskLoopDirective(S.getDirectiveKind())); |
| 6393 | // Emit outlined function for task construct. |
| 6394 | const CapturedStmt *CS = S.getCapturedStmt(OMPD_taskloop); |
| 6395 | Address CapturedStruct = Address::invalid(); |
| 6396 | { |
| 6397 | OMPLexicalScope Scope(*this, S, OMPD_taskloop, /*EmitPreInitStmt=*/false); |
| 6398 | CapturedStruct = GenerateCapturedStmtArgument(*CS); |
| 6399 | } |
| 6400 | QualType SharedsTy = getContext().getRecordType(CS->getCapturedRecordDecl()); |
| 6401 | const Expr *IfCond = nullptr; |
| 6402 | for (const auto *C : S.getClausesOfKind<OMPIfClause>()) { |
| 6403 | if (C->getNameModifier() == OMPD_unknown || |
| 6404 | C->getNameModifier() == OMPD_taskloop) { |
| 6405 | IfCond = C->getCondition(); |
| 6406 | break; |
| 6407 | } |
| 6408 | } |
| 6409 | |
| 6410 | OMPTaskDataTy Data; |
| 6411 | // Check if taskloop must be emitted without taskgroup. |
| 6412 | Data.Nogroup = S.getSingleClause<OMPNogroupClause>(); |
| 6413 | // TODO: Check if we should emit tied or untied task. |
| 6414 | Data.Tied = true; |
| 6415 | // Set scheduling for taskloop |
| 6416 | if (const auto* Clause = S.getSingleClause<OMPGrainsizeClause>()) { |
| 6417 | // grainsize clause |
| 6418 | Data.Schedule.setInt(/*IntVal=*/false); |
| 6419 | Data.Schedule.setPointer(EmitScalarExpr(Clause->getGrainsize())); |
| 6420 | } else if (const auto* Clause = S.getSingleClause<OMPNumTasksClause>()) { |
| 6421 | // num_tasks clause |
| 6422 | Data.Schedule.setInt(/*IntVal=*/true); |
| 6423 | Data.Schedule.setPointer(EmitScalarExpr(Clause->getNumTasks())); |
| 6424 | } |
| 6425 | |
| 6426 | auto &&BodyGen = [CS, &S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6427 | // if (PreCond) { |
| 6428 | // for (IV in 0..LastIteration) BODY; |
| 6429 | // <Final counter/linear vars updates>; |
| 6430 | // } |
| 6431 | // |
| 6432 | |
| 6433 | // Emit: if (PreCond) - begin. |
| 6434 | // If the condition constant folds and can be elided, avoid emitting the |
| 6435 | // whole loop. |
| 6436 | bool CondConstant; |
| 6437 | llvm::BasicBlock *ContBlock = nullptr; |
| 6438 | OMPLoopScope PreInitScope(CGF, S); |
| 6439 | if (CGF.ConstantFoldsToSimpleInteger(S.getPreCond(), CondConstant)) { |
| 6440 | if (!CondConstant) |
| 6441 | return; |
| 6442 | } else { |
| 6443 | llvm::BasicBlock *ThenBlock = CGF.createBasicBlock("taskloop.if.then" ); |
| 6444 | ContBlock = CGF.createBasicBlock("taskloop.if.end" ); |
| 6445 | emitPreCond(CGF, S, S.getPreCond(), ThenBlock, ContBlock, |
| 6446 | CGF.getProfileCount(&S)); |
| 6447 | CGF.EmitBlock(ThenBlock); |
| 6448 | CGF.incrementProfileCounter(&S); |
| 6449 | } |
| 6450 | |
| 6451 | (void)CGF.EmitOMPLinearClauseInit(S); |
| 6452 | |
| 6453 | OMPPrivateScope LoopScope(CGF); |
| 6454 | // Emit helper vars inits. |
| 6455 | enum { LowerBound = 5, UpperBound, Stride, LastIter }; |
| 6456 | auto *I = CS->getCapturedDecl()->param_begin(); |
| 6457 | auto *LBP = std::next(I, LowerBound); |
| 6458 | auto *UBP = std::next(I, UpperBound); |
| 6459 | auto *STP = std::next(I, Stride); |
| 6460 | auto *LIP = std::next(I, LastIter); |
| 6461 | mapParam(CGF, cast<DeclRefExpr>(S.getLowerBoundVariable()), *LBP, |
| 6462 | LoopScope); |
| 6463 | mapParam(CGF, cast<DeclRefExpr>(S.getUpperBoundVariable()), *UBP, |
| 6464 | LoopScope); |
| 6465 | mapParam(CGF, cast<DeclRefExpr>(S.getStrideVariable()), *STP, LoopScope); |
| 6466 | mapParam(CGF, cast<DeclRefExpr>(S.getIsLastIterVariable()), *LIP, |
| 6467 | LoopScope); |
| 6468 | CGF.EmitOMPPrivateLoopCounters(S, LoopScope); |
| 6469 | CGF.EmitOMPLinearClause(S, LoopScope); |
| 6470 | bool HasLastprivateClause = CGF.EmitOMPLastprivateClauseInit(S, LoopScope); |
| 6471 | (void)LoopScope.Privatize(); |
| 6472 | // Emit the loop iteration variable. |
| 6473 | const Expr *IVExpr = S.getIterationVariable(); |
| 6474 | const auto *IVDecl = cast<VarDecl>(cast<DeclRefExpr>(IVExpr)->getDecl()); |
| 6475 | CGF.EmitVarDecl(*IVDecl); |
| 6476 | CGF.EmitIgnoredExpr(S.getInit()); |
| 6477 | |
| 6478 | // Emit the iterations count variable. |
| 6479 | // If it is not a variable, Sema decided to calculate iterations count on |
| 6480 | // each iteration (e.g., it is foldable into a constant). |
| 6481 | if (const auto *LIExpr = dyn_cast<DeclRefExpr>(S.getLastIteration())) { |
| 6482 | CGF.EmitVarDecl(*cast<VarDecl>(LIExpr->getDecl())); |
| 6483 | // Emit calculation of the iterations count. |
| 6484 | CGF.EmitIgnoredExpr(S.getCalcLastIteration()); |
| 6485 | } |
| 6486 | |
| 6487 | { |
| 6488 | OMPLexicalScope Scope(CGF, S, OMPD_taskloop, /*EmitPreInitStmt=*/false); |
| 6489 | emitCommonSimdLoop( |
| 6490 | CGF, S, |
| 6491 | [&S](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6492 | if (isOpenMPSimdDirective(S.getDirectiveKind())) |
| 6493 | CGF.EmitOMPSimdInit(S); |
| 6494 | }, |
| 6495 | [&S, &LoopScope](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6496 | CGF.EmitOMPInnerLoop( |
| 6497 | S, LoopScope.requiresCleanups(), S.getCond(), S.getInc(), |
| 6498 | [&S](CodeGenFunction &CGF) { |
| 6499 | emitOMPLoopBodyWithStopPoint(CGF, S, |
| 6500 | CodeGenFunction::JumpDest()); |
| 6501 | }, |
| 6502 | [](CodeGenFunction &) {}); |
| 6503 | }); |
| 6504 | } |
| 6505 | // Emit: if (PreCond) - end. |
| 6506 | if (ContBlock) { |
| 6507 | CGF.EmitBranch(ContBlock); |
| 6508 | CGF.EmitBlock(ContBlock, true); |
| 6509 | } |
| 6510 | // Emit final copy of the lastprivate variables if IsLastIter != 0. |
| 6511 | if (HasLastprivateClause) { |
| 6512 | CGF.EmitOMPLastprivateClauseFinal( |
| 6513 | S, isOpenMPSimdDirective(S.getDirectiveKind()), |
| 6514 | CGF.Builder.CreateIsNotNull(CGF.EmitLoadOfScalar( |
| 6515 | CGF.GetAddrOfLocalVar(*LIP), /*Volatile=*/false, |
| 6516 | (*LIP)->getType(), S.getBeginLoc()))); |
| 6517 | } |
| 6518 | CGF.EmitOMPLinearClauseFinal(S, [LIP, &S](CodeGenFunction &CGF) { |
| 6519 | return CGF.Builder.CreateIsNotNull( |
| 6520 | CGF.EmitLoadOfScalar(CGF.GetAddrOfLocalVar(*LIP), /*Volatile=*/false, |
| 6521 | (*LIP)->getType(), S.getBeginLoc())); |
| 6522 | }); |
| 6523 | }; |
| 6524 | auto &&TaskGen = [&S, SharedsTy, CapturedStruct, |
| 6525 | IfCond](CodeGenFunction &CGF, llvm::Function *OutlinedFn, |
| 6526 | const OMPTaskDataTy &Data) { |
| 6527 | auto &&CodeGen = [&S, OutlinedFn, SharedsTy, CapturedStruct, IfCond, |
| 6528 | &Data](CodeGenFunction &CGF, PrePostActionTy &) { |
| 6529 | OMPLoopScope PreInitScope(CGF, S); |
| 6530 | CGF.CGM.getOpenMPRuntime().emitTaskLoopCall(CGF, S.getBeginLoc(), S, |
| 6531 | OutlinedFn, SharedsTy, |
| 6532 | CapturedStruct, IfCond, Data); |
| 6533 | }; |
| 6534 | CGF.CGM.getOpenMPRuntime().emitInlinedDirective(CGF, OMPD_taskloop, |
| 6535 | CodeGen); |
| 6536 | }; |
| 6537 | if (Data.Nogroup) { |
| 6538 | EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen, Data); |
| 6539 | } else { |
| 6540 | CGM.getOpenMPRuntime().emitTaskgroupRegion( |
| 6541 | *this, |
| 6542 | [&S, &BodyGen, &TaskGen, &Data](CodeGenFunction &CGF, |
| 6543 | PrePostActionTy &Action) { |
| 6544 | Action.Enter(CGF); |
| 6545 | CGF.EmitOMPTaskBasedDirective(S, OMPD_taskloop, BodyGen, TaskGen, |
| 6546 | Data); |
| 6547 | }, |
| 6548 | S.getBeginLoc()); |
| 6549 | } |
| 6550 | } |
| 6551 | |
| 6552 | void CodeGenFunction::EmitOMPTaskLoopDirective(const OMPTaskLoopDirective &S) { |
| 6553 | auto LPCRegion = |
| 6554 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6555 | EmitOMPTaskLoopBasedDirective(S); |
| 6556 | } |
| 6557 | |
| 6558 | void CodeGenFunction::EmitOMPTaskLoopSimdDirective( |
| 6559 | const OMPTaskLoopSimdDirective &S) { |
| 6560 | auto LPCRegion = |
| 6561 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6562 | OMPLexicalScope Scope(*this, S); |
| 6563 | EmitOMPTaskLoopBasedDirective(S); |
| 6564 | } |
| 6565 | |
| 6566 | void CodeGenFunction::EmitOMPMasterTaskLoopDirective( |
| 6567 | const OMPMasterTaskLoopDirective &S) { |
| 6568 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6569 | Action.Enter(CGF); |
| 6570 | EmitOMPTaskLoopBasedDirective(S); |
| 6571 | }; |
| 6572 | auto LPCRegion = |
| 6573 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6574 | OMPLexicalScope Scope(*this, S, llvm::None, /*EmitPreInitStmt=*/false); |
| 6575 | CGM.getOpenMPRuntime().emitMasterRegion(*this, CodeGen, S.getBeginLoc()); |
| 6576 | } |
| 6577 | |
| 6578 | void CodeGenFunction::EmitOMPMasterTaskLoopSimdDirective( |
| 6579 | const OMPMasterTaskLoopSimdDirective &S) { |
| 6580 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6581 | Action.Enter(CGF); |
| 6582 | EmitOMPTaskLoopBasedDirective(S); |
| 6583 | }; |
| 6584 | auto LPCRegion = |
| 6585 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6586 | OMPLexicalScope Scope(*this, S); |
| 6587 | CGM.getOpenMPRuntime().emitMasterRegion(*this, CodeGen, S.getBeginLoc()); |
| 6588 | } |
| 6589 | |
| 6590 | void CodeGenFunction::EmitOMPParallelMasterTaskLoopDirective( |
| 6591 | const OMPParallelMasterTaskLoopDirective &S) { |
| 6592 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6593 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 6594 | PrePostActionTy &Action) { |
| 6595 | Action.Enter(CGF); |
| 6596 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 6597 | }; |
| 6598 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 6599 | CGM.getOpenMPRuntime().emitMasterRegion(CGF, TaskLoopCodeGen, |
| 6600 | S.getBeginLoc()); |
| 6601 | }; |
| 6602 | auto LPCRegion = |
| 6603 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6604 | emitCommonOMPParallelDirective(*this, S, OMPD_master_taskloop, CodeGen, |
| 6605 | emitEmptyBoundParameters); |
| 6606 | } |
| 6607 | |
| 6608 | void CodeGenFunction::EmitOMPParallelMasterTaskLoopSimdDirective( |
| 6609 | const OMPParallelMasterTaskLoopSimdDirective &S) { |
| 6610 | auto &&CodeGen = [this, &S](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6611 | auto &&TaskLoopCodeGen = [&S](CodeGenFunction &CGF, |
| 6612 | PrePostActionTy &Action) { |
| 6613 | Action.Enter(CGF); |
| 6614 | CGF.EmitOMPTaskLoopBasedDirective(S); |
| 6615 | }; |
| 6616 | OMPLexicalScope Scope(CGF, S, OMPD_parallel, /*EmitPreInitStmt=*/false); |
| 6617 | CGM.getOpenMPRuntime().emitMasterRegion(CGF, TaskLoopCodeGen, |
| 6618 | S.getBeginLoc()); |
| 6619 | }; |
| 6620 | auto LPCRegion = |
| 6621 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, S); |
| 6622 | emitCommonOMPParallelDirective(*this, S, OMPD_master_taskloop_simd, CodeGen, |
| 6623 | emitEmptyBoundParameters); |
| 6624 | } |
| 6625 | |
| 6626 | // Generate the instructions for '#pragma omp target update' directive. |
| 6627 | void CodeGenFunction::EmitOMPTargetUpdateDirective( |
| 6628 | const OMPTargetUpdateDirective &S) { |
| 6629 | // If we don't have target devices, don't bother emitting the data mapping |
| 6630 | // code. |
| 6631 | if (CGM.getLangOpts().OMPTargetTriples.empty()) |
| 6632 | return; |
| 6633 | |
| 6634 | // Check if we have any if clause associated with the directive. |
| 6635 | const Expr *IfCond = nullptr; |
| 6636 | if (const auto *C = S.getSingleClause<OMPIfClause>()) |
| 6637 | IfCond = C->getCondition(); |
| 6638 | |
| 6639 | // Check if we have any device clause associated with the directive. |
| 6640 | const Expr *Device = nullptr; |
| 6641 | if (const auto *C = S.getSingleClause<OMPDeviceClause>()) |
| 6642 | Device = C->getDevice(); |
| 6643 | |
| 6644 | OMPLexicalScope Scope(*this, S, OMPD_task); |
| 6645 | CGM.getOpenMPRuntime().emitTargetDataStandAloneCall(*this, S, IfCond, Device); |
| 6646 | } |
| 6647 | |
| 6648 | void CodeGenFunction::EmitSimpleOMPExecutableDirective( |
| 6649 | const OMPExecutableDirective &D) { |
| 6650 | if (const auto *SD = dyn_cast<OMPScanDirective>(&D)) { |
| 6651 | EmitOMPScanDirective(*SD); |
| 6652 | return; |
| 6653 | } |
| 6654 | if (!D.hasAssociatedStmt() || !D.getAssociatedStmt()) |
| 6655 | return; |
| 6656 | auto &&CodeGen = [&D](CodeGenFunction &CGF, PrePostActionTy &Action) { |
| 6657 | OMPPrivateScope GlobalsScope(CGF); |
| 6658 | if (isOpenMPTaskingDirective(D.getDirectiveKind())) { |
| 6659 | // Capture global firstprivates to avoid crash. |
| 6660 | for (const auto *C : D.getClausesOfKind<OMPFirstprivateClause>()) { |
| 6661 | for (const Expr *Ref : C->varlists()) { |
| 6662 | const auto *DRE = cast<DeclRefExpr>(Ref->IgnoreParenImpCasts()); |
| 6663 | if (!DRE) |
| 6664 | continue; |
| 6665 | const auto *VD = dyn_cast<VarDecl>(DRE->getDecl()); |
| 6666 | if (!VD || VD->hasLocalStorage()) |
| 6667 | continue; |
| 6668 | if (!CGF.LocalDeclMap.count(VD)) { |
| 6669 | LValue GlobLVal = CGF.EmitLValue(Ref); |
| 6670 | GlobalsScope.addPrivate( |
| 6671 | VD, [&GlobLVal, &CGF]() { return GlobLVal.getAddress(CGF); }); |
| 6672 | } |
| 6673 | } |
| 6674 | } |
| 6675 | } |
| 6676 | if (isOpenMPSimdDirective(D.getDirectiveKind())) { |
| 6677 | (void)GlobalsScope.Privatize(); |
| 6678 | ParentLoopDirectiveForScanRegion ScanRegion(CGF, D); |
| 6679 | emitOMPSimdRegion(CGF, cast<OMPLoopDirective>(D), Action); |
| 6680 | } else { |
| 6681 | if (const auto *LD = dyn_cast<OMPLoopDirective>(&D)) { |
| 6682 | for (const Expr *E : LD->counters()) { |
| 6683 | const auto *VD = cast<VarDecl>(cast<DeclRefExpr>(E)->getDecl()); |
| 6684 | if (!VD->hasLocalStorage() && !CGF.LocalDeclMap.count(VD)) { |
| 6685 | LValue GlobLVal = CGF.EmitLValue(E); |
| 6686 | GlobalsScope.addPrivate( |
| 6687 | VD, [&GlobLVal, &CGF]() { return GlobLVal.getAddress(CGF); }); |
| 6688 | } |
| 6689 | if (isa<OMPCapturedExprDecl>(VD)) { |
| 6690 | // Emit only those that were not explicitly referenced in clauses. |
| 6691 | if (!CGF.LocalDeclMap.count(VD)) |
| 6692 | CGF.EmitVarDecl(*VD); |
| 6693 | } |
| 6694 | } |
| 6695 | for (const auto *C : D.getClausesOfKind<OMPOrderedClause>()) { |
| 6696 | if (!C->getNumForLoops()) |
| 6697 | continue; |
| 6698 | for (unsigned I = LD->getCollapsedNumber(), |
| 6699 | E = C->getLoopNumIterations().size(); |
| 6700 | I < E; ++I) { |
| 6701 | if (const auto *VD = dyn_cast<OMPCapturedExprDecl>( |
| 6702 | cast<DeclRefExpr>(C->getLoopCounter(I))->getDecl())) { |
| 6703 | // Emit only those that were not explicitly referenced in clauses. |
| 6704 | if (!CGF.LocalDeclMap.count(VD)) |
| 6705 | CGF.EmitVarDecl(*VD); |
| 6706 | } |
| 6707 | } |
| 6708 | } |
| 6709 | } |
| 6710 | (void)GlobalsScope.Privatize(); |
| 6711 | CGF.EmitStmt(D.getInnermostCapturedStmt()->getCapturedStmt()); |
| 6712 | } |
| 6713 | }; |
| 6714 | if (D.getDirectiveKind() == OMPD_atomic || |
| 6715 | D.getDirectiveKind() == OMPD_critical || |
| 6716 | D.getDirectiveKind() == OMPD_section || |
| 6717 | D.getDirectiveKind() == OMPD_master) { |
| 6718 | EmitStmt(D.getAssociatedStmt()); |
| 6719 | } else { |
| 6720 | auto LPCRegion = |
| 6721 | CGOpenMPRuntime::LastprivateConditionalRAII::disable(*this, D); |
| 6722 | OMPSimdLexicalScope Scope(*this, D); |
| 6723 | CGM.getOpenMPRuntime().emitInlinedDirective( |
| 6724 | *this, |
| 6725 | isOpenMPSimdDirective(D.getDirectiveKind()) ? OMPD_simd |
| 6726 | : D.getDirectiveKind(), |
| 6727 | CodeGen); |
| 6728 | } |
| 6729 | // Check for outer lastprivate conditional update. |
| 6730 | checkForLastprivateConditionalUpdate(*this, D); |
| 6731 | } |
| 6732 | |